BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Utah Center for Data Science//Talks//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALNAME:Utah Data Science & AI Lecture Series X-WR-CALDESC:Talks in the Data Science & AI Lecture Series at the Utah Cent er for Data Science. Generated from the talk records at datascience.utah.e du/talks/. X-WR-TIMEZONE:America/Denver REFRESH-INTERVAL;VALUE=DURATION:PT1H X-PUBLISHED-TTL:PT1H BEGIN:VTIMEZONE TZID:America/Denver BEGIN:DAYLIGHT TZOFFSETFROM:-0700 TZOFFSETTO:-0600 TZNAME:MDT DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0600 TZOFFSETTO:-0700 TZNAME:MST DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:2020-01-09-chris-musco@datascience.utah.edu DTSTAMP:20200109T191500Z DTSTART;TZID=America/Denver:20200109T121500 DTEND;TZID=America/Denver:20200109T133000 SUMMARY:Randomized FunctionalAnalysis - Chris Musco DESCRIPTION:Chris Musco (NYU)\n\nSketching and subsampling are central algo rithmic tools in scaling statisticalmethods to very large datasets. These techniques seek to quickly compress datadown to a compact set of informati ve features or examples\, which can then beprocessed in place of the origi nal data\, at much lower computational cost. Thecentral question of this t alk is what sketching methods can teach us abouteffective machine learning and data analysis in the small data regime. Inapplications where high qua lity data examples remain a rare luxury\, can ourknowledge of data sketchi ng guide more efficient initial data collection?\n\nWe study this problem by focusing specifically on techniques for large matrixcomputations. In th e field of randomized numerical linear algebra\, importancesampling has em erged as an important tool for dataset compression. Statisticalleverage sc ores and related measures are used to judge the importance of rowsor colum ns in a matrix\, which are then non-uniformly subsampled\, leading tofaste r algorithms for regression\, low-rank approximation\, kernel methods\, an dmany other data problems.\n\nI will introduce a simple generalization of leverage score sampling to infinitedimensional linear operators and show t he potential of this generalization indeveloping sample efficient algorith ms for small data applications.Specifically\, I will survey a number of re cent results on robust polynomialcurve fitting\, bandlimited function inte rpolation\, off-grid sparse Fouriertransforms\, and sample efficient covar iance estimation. I will illustrateconnections between these new results a nd classical tools in approximationtheory and signal processing\, and will discuss several open researchdirections.\n\nhttps://datascience.utah.edu/ talks/2020-01-09-chris-musco/ URL:https://datascience.utah.edu/talks/2020-01-09-chris-musco/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:algorithms & theory,machine learning,statistics END:VEVENT BEGIN:VEVENT UID:2020-01-16-alexander-lex@datascience.utah.edu DTSTAMP:20200116T191500Z DTSTART;TZID=America/Denver:20200116T121500 DTEND;TZID=America/Denver:20200116T133000 SUMMARY:Literate Visualization: Making Visual Analysis Sessions Reproducibl e and Reusable - Alexander Lex DESCRIPTION:Alexander Lex\n\nInteractive visualization is an important part of the data science process. It enables analysts to directly interact wit h the data\, exploring it with minimal effort. Unlike code\, however\, an interactive visualization session is ephemeral and can’t be easily share d\, revisited\, or reused. Computational notebooks\, such as Jupyter Noteb ooks\, R Markdown\, or Observable are a perfect match for many data scienc e applications. They are also the most popular embodiment of Knuth’s “ Literate Programming”\, where the logic of a program is explained in nat ural language\, figures\, and equations. In this talk\, I will sketch appr oaches to “Literate Visualization”. I will show how we can leverage pr ovenance data of an analysis session to create well-documented and annotat ed visualization stories that enable reproducibility and sharing. I will a lso introduce early work on semi-automatically inferring mid-level analysi s goals\, which allows us to understand the analysis process at a higher l evel. Understanding analysis goals enables us to speed up interactions and even re-used visual analysis processes.\n\nhttps://datascience.utah.edu/t alks/2020-01-16-alexander-lex/ URL:https://datascience.utah.edu/talks/2020-01-16-alexander-lex/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:visualization END:VEVENT BEGIN:VEVENT UID:2020-01-23-harish-maringanti@datascience.utah.edu DTSTAMP:20200123T191500Z DTSTART;TZID=America/Denver:20200123T121500 DTEND;TZID=America/Denver:20200123T133000 SUMMARY:Data Science projects in Marriott Library - Harish Maringanti DESCRIPTION:Harish Maringanti (Associate Dean for IT & Digital Library Serv ices\, Marriott Library)\n\nAt research intensive universities\, libraries have traditionally supported data science activities in various ways incl uding acquiring datasets that researchers need\, hosting workshops and tra ining sessions on data science tools\, and offering data support tools for creation of persistent identifiers(dois)\, etc. At Marriott Library\, in addition to supporting data science programs on campus\, we have embarked on a suite of data science projects to add value to our culturally-rich co llections. Our efforts are focused on enriching our collection data\, and making this collection data available for computational use (collections a s data [1]) so that developers\, scientists\, and digital humanists can pr ogrammatically interact with the data in myriad ways and undertake project s related to data mining & text analysis\, advanced visualizations\, and g eospatial analysis. In this presentation\, I will talk about two specific projects - Utah Digital Newspapers [2] and machine learning meets archives [3] - to highlight these efforts.\n\nUtah Digital Newspapers (UDN): Marri ott Library was an early pioneer in digitizing newspapers and making the c ontent available to historians\, researchers\, and lifelong learners. UDN program has been operating since 2002 and is recognized as one of the lead ers in newspaper digitization in the United States. We have continued to p artner with universities\, colleges\, state agencies\, county and city lib raries\, and other agencies to digitize\, deliver\, and archive historical newspaper collections\; As of 2019\, UDN has well over 22.5 million newsp aper articles and 3.5 million pages in the repository platform. In this pr esentation\, we will talk about the importance of looking at collections a s data\, our API work with UDN and demonstrate the usefulness of this appr oach with specific examples.\n\nMachine learning meets archives: Metadata is the bedrock of library archives and Digital Library systems\, as it hel ps in users discovering the unique content in various collections housed i n digital libraries. But creating metadata is a time-intensive process. We are working with machine learning algorithms to generate descriptive meta data for digital images. I will share the results of our work\, and also l essons learned from working with digital library data.\n\nhttps://datascie nce.utah.edu/talks/2020-01-23-harish-maringanti/ URL:https://datascience.utah.edu/talks/2020-01-23-harish-maringanti/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:machine learning END:VEVENT BEGIN:VEVENT UID:2020-01-30-qingyao-ai@datascience.utah.edu DTSTAMP:20200130T191500Z DTSTART;TZID=America/Denver:20200130T121500 DTEND;TZID=America/Denver:20200130T133000 SUMMARY:Unbiased Learning to Rank: Theory and Practice - Qingyao Ai DESCRIPTION:Qingyao Ai (Utah SoC)\n\nImplicit feedback (e.g.\, user clicks) is an important source of data for modern search engines. While heavily b iased\, it is cheap to collect and particularly useful for user-centric re trieval applications such as search ranking. Therefore\, a learning-to-ran k algorithm that can effectively learn from implicit user feedback without affected by its inherent biases could fundamentally change the design of ranking systems and significantly improve the quality of modern search eng ines. To develop an unbiased learning-to-rank system with biased feedback\ , previous studies have focused on constructing probabilistic graphical mo dels (e.g.\, click models) with user behavior hypothesis to extract and tr ain ranking systems with unbiased relevance signals. Recently\, a novel co unterfactual learning framework that estimates and adopts examination prop ensity for unbiased learning to rank has attracted much attention. In this talk\, we aim to provide an overview of the fundamental mechanism for unb iased learning to rank. We describe the theory behind existing frameworks\ , and give instructions on how to conduct unbiased learning to rank in pra ctice.\n\nhttps://datascience.utah.edu/talks/2020-01-30-qingyao-ai/ URL:https://datascience.utah.edu/talks/2020-01-30-qingyao-ai/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:fairness & ethics END:VEVENT BEGIN:VEVENT UID:2020-02-06-gail-zasowski@datascience.utah.edu DTSTAMP:20200206T191500Z DTSTART;TZID=America/Denver:20200206T121500 DTEND;TZID=America/Denver:20200206T133000 SUMMARY:Big Data\, Big Universe: Data-Driven Discoveries in Astrophysics - Gail Zasowski DESCRIPTION:Gail Zasowski (Utah Physics & Astronomy)\n\nThe stars in the ni ght sky have inspired questions about our place in the Universe throughout history. The development of telescopes showed us that the stars visible t o the naked eye are but a tiny fraction of their vast numbers within our o wn Galaxy\, and revealed energy signatures invisible to the human senses. We now know that there are billions of stars in our galaxy\, billions of g alaxies in our Universe\, and nearly 14 billion years of cosmic evolution that have led to where and what we are today. As the volume of astronomica l data grows at an ever quickening rate\, new discoveries increasingly com e from careful mining and analysis of existing data\, often used in unfore seen ways. This talk will describe some of the major unanswered questions in astrophysics\, and how new data-driven analysis techniques are uncoveri ng new insights into solving them.\n\nhttps://datascience.utah.edu/talks/2 020-02-06-gail-zasowski/ URL:https://datascience.utah.edu/talks/2020-02-06-gail-zasowski/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:physics & astronomy END:VEVENT BEGIN:VEVENT UID:2020-02-20-bei-wang@datascience.utah.edu DTSTAMP:20200220T191500Z DTSTART;TZID=America/Denver:20200220T121500 DTEND;TZID=America/Denver:20200220T133000 SUMMARY:TopoAct: Exploring the Shape of Activations in Deep Learning - Bei Wang DESCRIPTION:Bei Wang (Utah SoC\, SCI)\n\nDeep neural networks such as GoogL eNet and ResNet have achieved superhuman performance in tasks like image c lassification. To understand how such superior performance is achieved\, w e can probe a trained deep neural network by studying neuron activations\, that is\, combinations of neuron firings\, at any layer of the network in response to a particular input. With a large set of input images\, we aim to obtain a global view of what neurons detect by studying their activati ons. We ask the following questions: What is the shape of the space of act ivations? That is\, what is the organizational principle behind neuron act ivations\, and how are the activations related within a layer and across l ayers? Applying tools from topological data analysis\, we present TopoAct\ , a visual exploration system used to study topological summaries of activ ation vectors for a single layer as well as the evolution of such summarie s across multiple layers. We present visual exploration scenarios using To poAct that provide valuable insights towards learned representations of an image classifier.\n\nhttps://datascience.utah.edu/talks/2020-02-20-bei-wa ng/ URL:https://datascience.utah.edu/talks/2020-02-20-bei-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:deep learning,machine learning END:VEVENT BEGIN:VEVENT UID:2020-02-27-taylor-sparks@datascience.utah.edu DTSTAMP:20200227T191500Z DTSTART;TZID=America/Denver:20200227T121500 DTEND;TZID=America/Denver:20200227T133000 SUMMARY:New Algorithms\, Descriptors\, and Machine Learning Techniques Tail ored to the Challenges of Materials Informatics - Taylor Sparks DESCRIPTION:Taylor Sparks (Utah Materials Science & Engineering)\n\nNew mat erials are required to address many of the energy\, environmental\, and te chnological needs of the present and future. Materials Informatics\, or th e application of data science techniques to solve materials research chall enges\, is expected to play a key role in materials development and discov ery given the infinite palette available for new materials. Interestingly\ , the requirements and tasks of Materials Informatics do not always overla p with general machine learning. Therefore\, adopting existing data scienc e tools including visualization\, algorithms\, featurization schemas etc m ay not provide the ideal outcomes for the specific needs of Materials Info rmatics.\n\nIn this talk I will focus on some of our recent work to bring tailored data science approaches to actual materials research problems. Sp ecifically\, I will introduce how we have created an attention-based neura l network architecture for the prediction of materials properties. We show that this novel algorithm outperforms other methods in the absence of che mical information\, even when the statistical and ensemble learning techni ques are given domain-specific chemical knowledge about the materials.\n\n Dr. Sparks is an Associate Professor and Associate Chair of the Materials Science and Engineering Department at the University of Utah. He is origin ally from Utah and an alumni of the department he now teaches in. Before g raduate school he worked at Ceramatec Inc. He did his MS in Materials at U CSB and his PhD in Applied Physics at Harvard University in David Clarke ’s laboratory and then did a postdoc with Ram Seshadri in the Materials Research Laboratory at UCSB. His current research centers on the discovery \, synthesis\, characterization\, and properties of new materials for ener gy applications. He is a pioneer in the emerging field of materials inform atics whereby big data\, data mining\, and machine learning are leveraged to solve challenges in materials science. He also hosts a podcast entitled “Materialism” where he discusses the past\, present\, and future of M aterials Science.\n\nhttps://datascience.utah.edu/talks/2020-02-27-taylor- sparks/ URL:https://datascience.utah.edu/talks/2020-02-27-taylor-sparks/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:algorithms & theory,machine learning,physics & astronomy END:VEVENT BEGIN:VEVENT UID:2020-03-05-john-horel@datascience.utah.edu DTSTAMP:20200305T191500Z DTSTART;TZID=America/Denver:20200305T121500 DTEND;TZID=America/Denver:20200305T133000 SUMMARY:The Promise and Perils of Big Data in the Cloud - Examples from the Atmospheric Sciences - John Horel DESCRIPTION:John Horel (Professor\, Chair\, Department of Atmospheric Scien ces)\n\nFrom the inception of numerical weather prediction in the 1950’s \, atmospheric scientists have stretched the envelope on the hardware and procedures available to write\, store\, and use data on mass storage syste ms. The opportunities now to rely on cloud resources to process\, access\, and disseminate environmental data offer improved capabilities for data s cience applications in the atmospheric sciences but also introduce complex ities for university researchers.\n\nThe Big Data Project of the National Oceanographic and Atmospheric Administration is assessing the potential be nefits of storing in the cloud observations and weather and climate model output that are generating petabytes of data daily. Retrieving\, archiving \, analyzing\, and disseminating only a fraction of this environmental inf ormation has required us to move beyond computational approaches tradition ally used within the atmospheric science community. For example\, hundreds of users rely on a 140+ Tbyte archive we maintain of High Resolution Rapi d Refresh (HRRR) model output on the Pando system of the University’s Ce nter for High Performance Computing. Computing resources available nationw ide as part of the Open Science Grid- a high-throughput computing resource - have been used to analyze wildland fire events. Data analytical methods are being explored that rely on Zarr data compression to provide classes a nd functions for working with N-dimensional arrays.\n\nhttps://datascience .utah.edu/talks/2020-03-05-john-horel/ URL:https://datascience.utah.edu/talks/2020-03-05-john-horel/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:climate & environment END:VEVENT BEGIN:VEVENT UID:2020-08-28-vivek-gupta@datascience.utah.edu DTSTAMP:20200828T175000Z DTSTART;TZID=America/Denver:20200828T115000 DTEND;TZID=America/Denver:20200828T131000 SUMMARY:InfoTabS: Inference on Tables as Semi-Structured Data - Vivek Gupta DESCRIPTION:Vivek Gupta (UoU)\n\nExperience of the everyday language indica tes the use of complicated reasonings both for people and the AI systems. Natural Language Inference (NLI) is the process of reasoning about inferen tial relationships\, meaning to establish whether a hypothesis is a true ( entailment)\, false (contradiction)\, or undetermined (neutral) given a pr emise. Previous works have generated inference corpora\, such as the SNLI and the MNLI\, which comprise only unstructured representations of text in the form of sentences in which relationships between words are explicitly expressed\, and often need information extraction. However\, text can als o occur universally in other structured forms like tables\, graphs\, and d atabases.Building upon previous work on large-scale datasets for inference \, we introduce a new dataset called INFOTABS\, comprising of human-writte n textual hypotheses based on premises that are tables extracted from Wiki pedia info-boxes. Our analysis shows that the semi-structured\, multi-doma in\, and heterogeneous nature of the premises admits complex\, multi-facet ed reasoning. Experiments reveal that\, while human annotators agree on th e relationships between a table-hypothesis pair\, several standard modelin g strategies are unsuccessful at the task\, suggesting that reasoning abou t tables can pose a new modeling challenge. For more details on InfoTabS v isit http://infotabs.github.io (http://infotabs.github.io/)\n\nRecording: https://www.youtube.com/redirect?q=https%3A%2F%2Fvgupta123.github.io%2F&v= YhfU1BON8EI&event=video_description&redir_token=QUFFLUhqbFVhUUxOS2RDSHBPYW JPMjNqaFpoalpzY2E3d3xBQ3Jtc0tuWlZRQ0tfSTdPNkxsTzFuWDN2Y3lvdXdwRXJrOTJ6bGI3 NXhrc2x4QVYxalNTTUVXWE10V3VncklsZ2Q1RnltN1VaWkpsek9OY0EtdWwtQWZNVHR0Q2k4Rn g3eVE4N3FWYTk5dUtETV9aZXRSaGlEcw%3D%3D\n\nhttps://datascience.utah.edu/tal ks/2020-08-28-vivek-gupta/ URL:https://datascience.utah.edu/talks/2020-08-28-vivek-gupta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:data management,education,machine learning,natural language proc essing END:VEVENT BEGIN:VEVENT UID:2020-09-04-dheeraj-mekala@datascience.utah.edu DTSTAMP:20200904T175000Z DTSTART;TZID=America/Denver:20200904T115000 DTEND;TZID=America/Denver:20200904T131000 SUMMARY:Contextualized Weak Supervision for Text Classification - Dheeraj M ekala DESCRIPTION:Dheeraj Mekala (UCSD)\n\nWeakly supervised text classification based on a few user-provided seed words has recently attracted much attent ion from researchers. Existing methods mainly generate pseudo-labels in a context-free manner (e.g.\, string matching)\, therefore\, the ambiguous\, context-dependent nature of human language has been long overlooked. In t his paper\, we propose a novel framework ConWea\, providing contextualized weak supervision for text classification. Specifically\, we leverage cont extualized representations of word occurrences and seed word information t o automatically differentiate multiple interpretations of the same word\, and thus create a contextualized corpus. This contextualized corpus is fur ther utilized to train the classifier and expand seed words in an iterativ e manner. This process not only adds new contextualized\, highly label-ind icative keywords but also disambiguates initial seed words\, making our we ak supervision fully contextualized. Extensive experiments and case studie s on real-world datasets demonstrate the necessity and significant advanta ges of using contextualized weak supervision\, especially when the class l abels are fine-grained.\n\nhttps://datascience.utah.edu/talks/2020-09-04-d heeraj-mekala/ URL:https://datascience.utah.edu/talks/2020-09-04-dheeraj-mekala/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:machine learning,natural language processing END:VEVENT BEGIN:VEVENT UID:2020-09-11-parthe-pandit@datascience.utah.edu DTSTAMP:20200911T175000Z DTSTART;TZID=America/Denver:20200911T115000 DTEND;TZID=America/Denver:20200911T131000 SUMMARY:Characterizing the asymptotic performance of inverse problems over Deep Networks - Parthe Pandit DESCRIPTION:Parthe Pandit (UCLA)\n\nAt the heart of Machine Learning lies t he question of generalizability of learned models over previously unseen d ata. While over-parameterized models based on neural networks are now ubiq uitous in machine learning applications\, our understanding of their gener alization capabilities remains incomplete. This task is made harder by the non-convexity of the underlying estimation/learning problems.\n\nThe solu tions to some of these estimation problems can be analyzed using a class o f algorithms called Approximate Message Passing\, even in the presence of the non-convexity. The dynamics of this algorithm follow a simplified macr oscopic description often called the State Evolution. This analytical tool allows us to provide some insights into two broad classes of problems rel ated to Neural Networks:\n\n1. Generalization error in 1 and 2-layer Netwo rks\n2. Image Reconstruction error with Deep Image Priors\n\nhttps://datas cience.utah.edu/talks/2020-09-11-parthe-pandit/ URL:https://datascience.utah.edu/talks/2020-09-11-parthe-pandit/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:algorithms & theory,deep learning,machine learning,optimization END:VEVENT BEGIN:VEVENT UID:2020-09-18-vishnu-lokhande@datascience.utah.edu DTSTAMP:20200918T175000Z DTSTART;TZID=America/Denver:20200918T115000 DTEND;TZID=America/Denver:20200918T131000 SUMMARY:Optimization methods for imposing Fairness in Computer Vision Model s - Vishnu Lokhande DESCRIPTION:Vishnu Lokhande (University of Wisconsin-Madison)\n\nIn this ta lk\, we will study a mechanism to impose fairness in computer vision model s concurrently while training the model and informed by standard fairness measures. While existing fairness based approaches in vision have largely relied on training adversarial modules together with the primary classific ation/regression task\, in an effort to remove the influence of the protec ted attribute or variable\, we will discuss how ideas based on well-known optimization concepts can provide a simpler alternative. In our proposed s cheme\, imposing fairness just requires specifying the protected attribute and utilizing our optimization routine. We will discuss experiments\, tha t are interpretable\, demonstrating that several fairness measures from th e literature can be reliably imposed on standard vision tasks. We will als o discuss technical analysis on the convergence guarantees of the said opt imization routine.\n\nhttps://datascience.utah.edu/talks/2020-09-18-vishnu -lokhande/ URL:https://datascience.utah.edu/talks/2020-09-18-vishnu-lokhande/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:computer vision,fairness & ethics,machine learning,optimization END:VEVENT BEGIN:VEVENT UID:2020-09-25-matthew-h-samore-jeffrey-humpherys@datascience.utah.edu DTSTAMP:20200925T175000Z DTSTART;TZID=America/Denver:20200925T115000 DTEND;TZID=America/Denver:20200925T131000 SUMMARY:Learning to care - Matthew H. Samore\, Jeffrey Humpherys DESCRIPTION:Matthew H. Samore (Utah Epidemiology)\, Jeffrey Humpherys (Utah Epidemiology)\n\nhttps://datascience.utah.edu/talks/2020-09-25-matthew-h- samore-jeffrey-humpherys/ URL:https://datascience.utah.edu/talks/2020-09-25-matthew-h-samore-jeffrey- humpherys/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 END:VEVENT BEGIN:VEVENT UID:2020-10-02-ellen-riloff@datascience.utah.edu DTSTAMP:20201002T175000Z DTSTART;TZID=America/Denver:20201002T115000 DTEND;TZID=America/Denver:20201002T131000 SUMMARY:Identifying Affective Events and the Reasons for their Polarity - E llen Riloff DESCRIPTION:Ellen Riloff (Utah Computer Science)\n\nRecognizing affective s tates is essential for narrative text\nunderstanding and for applications such as conversational dialogue\,\nsummarization\, and sarcasm recognition . Many tools have been developed\nto recognize explicit expressions of sen timent\, but affective states\ncan also be inferred from events. This talk will focus on "affective\nevents"\, which are generally desirable or unde sirable experiences that\nimplicitly suggest an affective state for the ex periencer. For\nexample\, buying a home is usually desirable and associate d with a\npositive affective state\, but being laid off is undesirable and \nassociated with a negative state. First\, we will describe a weakly\nsup ervised learning method to induce affective events from a text\ncorpus by optimizing for semantic consistency. Second\, we aim to\ncharacterize affe ctive events based on Human Needs Categories\, which\noften explain people 's motivations\, goals\, and desires. We will\npresent a co-training model for Human Needs categorization that uses\nan event expression classifier and an event context classifier to\nlearn from both labeled and unlabeled texts.\n\nhttps://datascience.utah.edu/talks/2020-10-02-ellen-riloff/ URL:https://datascience.utah.edu/talks/2020-10-02-ellen-riloff/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:machine learning,natural language processing END:VEVENT BEGIN:VEVENT UID:2020-10-09-swaroop-mishra@datascience.utah.edu DTSTAMP:20201009T175000Z DTSTART;TZID=America/Denver:20201009T115000 DTEND;TZID=America/Denver:20201009T131000 SUMMARY:DQI: Measuring Data Quality in NLP - Swaroop Mishra DESCRIPTION:Swaroop Mishra (Arizona State University)\n\nNeural language mo dels have achieved human-level performance across several NLP datasets. Ho wever\, recent studies have shown that these models are not truly learning the desired task\; rather\, their high performance is attributed to overf itting using spurious biases\, which suggests that the capabilities of AI systems have been over-estimated. We introduce a generic formula for Data Quality Index (DQI) to help dataset creators create datasets with minimal unwanted biases. We propose a new data creation paradigm using DQI to crea te higher quality data. The data creation paradigm consists of several dat a visualizations to help data creators (i) understand the quality of data and (ii) visualize the impact of the created data instance on the overall quality. It also has a couple of automation methods to (i) assist data cre ators and (ii) make the model more robust to adversarial attacks. We use D QI along with these automation methods to renovate biased examples in SNLI . We show that models trained on the renovated SNLI dataset generalize bet ter to out of distribution tasks. Renovation results in reduced model perf ormance\, exposing a large gap with respect to human performance. DQI syst ematically helps in creating harder benchmarks using active learning. Our work takes the process of dynamic dataset creation forward\, wherein datas ets evolve together with the evolving state of the art\, therefore serving as a means of benchmarking the true progress of AI. Finally\, we also sho w that DQI helps in pruning a dataset without compromising IID and OOD per formance significantly.\n\nhttps://datascience.utah.edu/talks/2020-10-09-s waroop-mishra/ URL:https://datascience.utah.edu/talks/2020-10-09-swaroop-mishra/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:data management,fairness & ethics,natural language processing END:VEVENT BEGIN:VEVENT UID:2020-10-16-varun-gangal@datascience.utah.edu DTSTAMP:20201016T175000Z DTSTART;TZID=America/Denver:20201016T115000 DTEND;TZID=America/Denver:20201016T131000 SUMMARY:Examining Extra Sentential Abilities of Contextual Embeddings - Var un Gangal DESCRIPTION:Varun Gangal (LTI\, CMU)\n\nIn the first third of our talk\, we try to understand what and how much does BERT already know about event ar guments (including cross-sentence ones)?. We observe that BERT's attention heads have modest but well above-chance ability to spot event arguments s ans any training. Furthermore\, we investigate how our methods do for cros s-sentence event arguments\, proposing a procedure to isolate "best heads" for cross-sentence argument detection separately of those for intra-sente nce arguments. In the second third\, we take a closer look at the infillin g abilities of BERT. We know BERT is good at Word-level Infilling (obvious ly!) . How good is it though at Sentence-level Infilling a.k.a Cloze ? We introduce a human-created sentence cloze dataset\, collected from public s chool English examinations. Our task requires a model to fill up multiple blanks in a passage from a shared candidate set with distractors designed by English teachers. Our experiments show a significant performance gap be tween BERT (72%) and humans (87%)\, encouraging future models to bridge th is gap.We conclude our talk by going through a somewhat unrelated\, recent foray into data augmentation for finetuning pretrained generators on low resource domains.\n\nhttps://datascience.utah.edu/talks/2020-10-16-varun-g angal/ URL:https://datascience.utah.edu/talks/2020-10-16-varun-gangal/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:deep learning END:VEVENT BEGIN:VEVENT UID:2020-10-23-nancy-wang@datascience.utah.edu DTSTAMP:20201023T175000Z DTSTART;TZID=America/Denver:20201023T115000 DTEND;TZID=America/Denver:20201023T131000 SUMMARY:Global Table Extractor (GTE): A Framework for Joint Table Identific ation and Cell Structure Recognition Using Visual Context - Nancy Wang DESCRIPTION:Nancy Wang (IBM)\n\nDocuments are often the format of choice fo r knowledge sharing and preservation in business and science\, within whic h are tables that capture most of the critical data. Unfortunately\, most documents are stored and distributed as PDF or scanned images\, which fail to preserve table formatting.\nRecent vision-based deep learning approach es have been proposed to address this gap\, but most still cannot achieve state-of-the-art results.\n\n We present Global Table Extractor (GTE)\, a vision-guided systematic framework for joint table detection and cell stru ctured recognition\, which could be built on top of any object detection m odel. With GTE-Table\, we invent a new penalty based on the natural cell c ontainment constraint of tables to train our table network aided by cell l ocation predictions. GTE-Cell is a new hierarchical cell detection network that leverages table styles. Further\, we design a method to automaticall y label table and cell structure in existing documents to cheaply create a large corpus of training and test data. We use this to enhance PubTabNet with cell labels and create FinTabNet\, real-world and complex scientific and financial datasets with detailed table structure annotations to help t rain and test structure recognition.\n\n Our deep learning framework surpa sses previous state-of-the-art results on the ICDAR 2013 and ICDAR 2019 ta ble competition test dataset in both table detection and cell structure re cognition. Further experiments demonstrate a greater than 45% improvement in cell structure recognition when compared to a vanilla RetinaNet object detection model in our new financial dataset (FinTabNet).\n\nhttps://datas cience.utah.edu/talks/2020-10-23-nancy-wang/ URL:https://datascience.utah.edu/talks/2020-10-23-nancy-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:biology & genomics,computer vision,data management,deep learning END:VEVENT BEGIN:VEVENT UID:2020-10-30-daniel-scharfstein@datascience.utah.edu DTSTAMP:20201030T175000Z DTSTART;TZID=America/Denver:20201030T115000 DTEND;TZID=America/Denver:20201030T131000 SUMMARY:Semiparametrics: A Biostatistician’s Toolbox - Daniel Scharfstein DESCRIPTION:Daniel Scharfstein (Utah\, Population Health Sciences)\n\nIn th is talk\, I will discuss the theory of semiparametrics that I use to estim ate causal effects at root-n rates. Estimators of these effects depend on estimators of nuisance parameters that can be estimated at rates slower th an root-n\; I provide sufficient conditions for these rates. I will seek a dvice on the machine learning estimation techniques that satisfy these con ditions. I will illustrate the theory in the context of estimating the cau sal contrast of two competing treatments based on data from a comprehensiv e cohort study in which clinically eligible individuals are first asked t o enroll in a randomized trial and\, if they refuse\, are then asked to en roll in a parallel observational study in which they can choose treatment according to their own preference.\n\nhttps://datascience.utah.edu/talks/2 020-10-30-daniel-scharfstein/ URL:https://datascience.utah.edu/talks/2020-10-30-daniel-scharfstein/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:causal inference,health & medicine,statistics END:VEVENT BEGIN:VEVENT UID:2020-10-30-alberto-cairo@datascience.utah.edu DTSTAMP:20201030T213000Z DTSTART;TZID=America/Denver:20201030T153000 DTEND;TZID=America/Denver:20201030T163000 SUMMARY:Data Visualization: How to Make Good Decisions - Alberto Cairo DESCRIPTION:Alberto Cairo\n\nData visualization\, the display of data throu gh graphs\, charts\, maps\, and diagrams\, is a skill in great demand in m any disciplines\, from the sciences to communication or business analytics . However\, visualization is often misunderstood. For instance\, it's ofte n taught as the application of a series of strict rules. This talk argues that visualization is more akin to writing: yes\, we do need to understand visualization's grammar but\, beyond that\, visualization design is flexi ble\, and can't be based on rules that are set in stone. Instead\, designe rs need to develop a good decision-making framework based on asking themse lves a series of questions.\n\nhttps://datascience.utah.edu/talks/2020-10- 30-alberto-cairo/ URL:https://datascience.utah.edu/talks/2020-10-30-alberto-cairo/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE CATEGORIES:human-centered computing,networks & graphs,visualization END:VEVENT BEGIN:VEVENT UID:2020-11-06-bhargavi-paranjape@datascience.utah.edu DTSTAMP:20201106T185000Z DTSTART;TZID=America/Denver:20201106T115000 DTEND;TZID=America/Denver:20201106T131000 SUMMARY:An Information Bottleneck Approach for Rationale Extraction - Bharg avi Paranjape DESCRIPTION:Bhargavi Paranjape (University of Washington)\n\nDecisions of c omplex models for language understanding can be explained by limiting the inputs they are provided to a relevant sub-sequence of the original text — a rationale. Models that condition predictions on a concise rationale\ , while being more interpretable\, tend to be less accurate than models th at are able to use the entire context. In this paper\, we show that it is possible to better manage the trade-off between concise explanations and h igh task accuracy by optimizing abound on the Information Bottleneck (IB) objective. Our approach jointly learns an explainer that predicts sparse b inary masks over input sentences without explicit supervision and an end-t ask predictor that considers only the residual sentences. Using IB\, we de rive a learning objective that allows direct control of mask sparsity leve ls through a tunable sparse prior. Experiments on the ERASER benchmark dem onstrate significant gains over previous work for both task performance an d agreement with human rationales\n\nhttps://datascience.utah.edu/talks/20 20-11-06-bhargavi-paranjape/ URL:https://datascience.utah.edu/talks/2020-11-06-bhargavi-paranjape/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:education,fairness & ethics,natural language processing END:VEVENT BEGIN:VEVENT UID:2020-11-13-akanksha-atrey@datascience.utah.edu DTSTAMP:20201113T185000Z DTSTART;TZID=America/Denver:20201113T115000 DTEND;TZID=America/Denver:20201113T131000 SUMMARY:Towards High-Performance Machine Learning on the Edge - Akanksha At rey DESCRIPTION:Akanksha Atrey (UMass)\n\nModern day distributed technologies\, such as mobile systems and the Internet of Things (IoT)\, enable the glob al integration of heterogeneous smart devices via wireless networks. A com mon characteristic across these technologies is their ability to collect a nd communicate continuously streaming data. The generation of high bandwid th data makes machine learning (ML) and artificial intelligence (AI) appea ling for processing\, reasoning\, and predicting about the environment\, b ut low network latency requirements make offloading intelligence to the cl oud undesirable. This raises an important question: how can we design\, de velop and evaluate ML algorithms that perform beyond predictive accuracy ( e.g.\, generalizable\, explainable\, privacy-aware\, and efficient) while being accessible and scalable in resource-constrained edge environments? I n this talk\, I will cover two aspects\, explainability and privacy\, when deploying such ML models in modern distributed technologies. The focus of the talk will be two-fold: (1) counterfactual evaluation of the explanati ons generated using saliency maps in deep reinforcement learning for appli cations such as autonomous vehicles\, and (2) privacy implications of pers onalized ML models in context-aware mobility applications. The talk will b e concluded with a discussion on where the future of ML lies in evolving d istributed technologies.\n\nhttps://datascience.utah.edu/talks/2020-11-13- akanksha-atrey/ URL:https://datascience.utah.edu/talks/2020-11-13-akanksha-atrey/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:algorithms & theory,machine learning,privacy & security END:VEVENT BEGIN:VEVENT UID:2020-11-20-akhil-arora@datascience.utah.edu DTSTAMP:20201120T185000Z DTSTART;TZID=America/Denver:20201120T115000 DTEND;TZID=America/Denver:20201120T131000 SUMMARY:Low-rank Subspaces for Unsupervised Entity Linking - Akhil Arora DESCRIPTION:Akhil Arora (EPFL)\n\nEntity linking is an important problem wi th many applications. Most previous solutions were designed for settings w here annotated training data is available\, which is\, however\, not the c ase in numerous domains. We propose a light-weight and scalable entity lin king method\, Eigenthemes\, that relies solely on the availability of enti ty names and a referent knowledge base. Eigenthemes exploits the fact that the entities that are truly mentioned in a document (the ``gold entities' ') tend to form a semantically dense subset of the set of all candidate en tities in the document. Geometrically speaking\, when representing entitie s as vectors via some given embedding\, the gold entities tend to lie in a low-rank subspace of the full embedding space. Eigenthemes identifies thi s subspace using the singular value decomposition and scores candidate ent ities according to their proximity to the subspace. Extensive experiments on benchmark datasets from a variety of real-world domains showcase the ef fectiveness of our approach.\n\nhttps://datascience.utah.edu/talks/2020-11 -20-akhil-arora/ URL:https://datascience.utah.edu/talks/2020-11-20-akhil-arora/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:deep learning,machine learning END:VEVENT BEGIN:VEVENT UID:2020-12-04-danish-pruthi@datascience.utah.edu DTSTAMP:20201204T185000Z DTSTART;TZID=America/Denver:20201204T115000 DTEND;TZID=America/Denver:20201204T131000 SUMMARY:A Tale of Evidence and Explanations - Danish Pruthi DESCRIPTION:Danish Pruthi (LTI\, CMU)\n\nI would present a brief overview o f the state of research in explainability and its evaluation (or lack ther eof). Then\, I would offer a new lens into explanations\, viewing them as a communication channel between a teacher and a student. This view enables us to quantitatively evaluate different attribution methods in a principl ed way at scale. Shifting gears\, in the second part of the talk\, I would introduce new techniques to supplement predictions with evidence to enabl e stakeholders to verify the outcomes readily.\n\nhttps://datascience.utah .edu/talks/2020-12-04-danish-pruthi/ URL:https://datascience.utah.edu/talks/2020-12-04-danish-pruthi/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/83503994251?pwd=Wm01bk40UWxva2pSV0dTbFZH bGtLQT09 CATEGORIES:education END:VEVENT BEGIN:VEVENT UID:2021-01-22-grad-student-spotlights@datascience.utah.edu DTSTAMP:20210122T210000Z DTSTART;TZID=America/Denver:20210122T140000 DTEND;TZID=America/Denver:20210122T150000 SUMMARY:1. Introduction and Logistics - Grad Student Spotlights DESCRIPTION:Grad Student Spotlights (10-minute talks by 4 current graduate students)\n\nArchit Rathore: Exploring the Shape of Activations - A BERT c ase studyBenwei Shi: At-the-time and Back-in-time Persistent Sketches\nJoe Vinu: What all it takes for Performant Deep Nets to be Reliable and Fair? \nBrian Lavallee: Rounding Out Structural Rounding\nVivek Gupta: Inference on Tables as Semi-Structured Data\n\nhttps://datascience.utah.edu/talks/2 021-01-22-grad-student-spotlights/ URL:https://datascience.utah.edu/talks/2021-01-22-grad-student-spotlights/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:data management,fairness & ethics END:VEVENT BEGIN:VEVENT UID:2021-01-29-sanghamitra-dutta@datascience.utah.edu DTSTAMP:20210129T210000Z DTSTART;TZID=America/Denver:20210129T140000 DTEND;TZID=America/Denver:20210129T150000 SUMMARY:A Systematic Understanding of Exempt and Non-Exempt Algorithmic Bia ses - Sanghamitra Dutta DESCRIPTION:Sanghamitra Dutta (CMU)\n\nWith the growing use of machine lear ning algorithms in highly consequential domains\, the quantification and r emoval of bias with respect to gender\, race\, etc.\, is becoming increasi ngly important. While quantifying bias is essential\, sometimes the needs of a business (e.g.\, hiring) may require the use of certain features that are critical in a way that any bias that can be explained by them might n eed to be exempted (inspired from the business necessity defense of Title VII of Civil Rights Act). For instance\, in hiring a software engineer\, a standardized coding-test score may be a critical feature that is weighed strongly in the decision even if it introduces bias\, whereas other featur es\, such as name\, zip code\, or reference letters may be used to improve decision-making\, but only to the extent that they do not introduce bias. In this work\, we propose a novel information-theoretic measure of non-ex empt bias\, which quantifies the part of the bias that cannot be accounted for by the critical features. This measure can be applied for (i) Auditin g trained models to check if the bias arose purely due to the critical fea tures\; and also for (ii) Training with selective removal of the non-exemp t bias if desired. We arrive at this decomposition through canonical examp les that lead to a set of desirable properties (axioms) that any measure o f non-exempt bias should satisfy. We then propose a causal measure of non- exempt bias that satisfies all of them. We also propose observational meas ures that only satisfy some of these properties (including an impossibilit y result on observational measures being able to satisfy all properties). Then\, we perform case studies using them to show how one can train models while reducing non-exempt bias. Our quantification bridges ideas of causa lity\, Simpson's paradox\, and a body of work from information theory call ed Partial Information Decomposition (PID). The talk will be fairly access ible\, no knowledge of information theory is required.\n\nhttps://datascie nce.utah.edu/talks/2021-01-29-sanghamitra-dutta/ URL:https://datascience.utah.edu/talks/2021-01-29-sanghamitra-dutta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:algorithms & theory,causal inference,fairness & ethics,machine l earning END:VEVENT BEGIN:VEVENT UID:2021-02-05-michal-moshkovitz@datascience.utah.edu DTSTAMP:20210205T210000Z DTSTART;TZID=America/Denver:20210205T140000 DTEND;TZID=America/Denver:20210205T150000 SUMMARY:Unexpected Effects of Online no-Substitution k-means Clustering - M ichal Moshkovitz DESCRIPTION:Michal Moshkovitz (UCSD)\n\nOffline k-means clustering was stud ied extensively and algorithms with a constant approximation are available . However\, online clustering is still uncharted. New factors come into pl ay: the ordering of the dataset and whether the number of points\, n\, is known in advance or not. Their exact effects are unknown. In this work\, w e focus on the online setting where the decisions are irreversible: after a point arrives the algorithm needs to decide whether to take the point as a center or not\, and this decision is final. How many centers are needed and sufficient to achieve constant approximation in this setting? We show upper and lower bounds for all the different cases. These bounds are exac tly the same up to a constant\, thus achieving optimal bounds. For example \, for k-means cost with constant k>1 and random order\, Θ(logn) centers are enough to achieve a constant approximation\, while the mere a priori k nowledge of n reduces the number of centers to a constant. These bounds ho ld for any distance function that obeys a triangle-type inequality.\n\nhtt ps://datascience.utah.edu/talks/2021-02-05-michal-moshkovitz/ URL:https://datascience.utah.edu/talks/2021-02-05-michal-moshkovitz/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:algorithms & theory END:VEVENT BEGIN:VEVENT UID:2021-02-12-fritz-lekschas@datascience.utah.edu DTSTAMP:20210212T210000Z DTSTART;TZID=America/Denver:20210212T140000 DTEND;TZID=America/Denver:20210212T150000 SUMMARY:Visual Pattern Exploration At and Across Scales - Fritz Lekschas DESCRIPTION:Fritz Lekschas (Harvard)\n\nVisually exploring data is a powerf ul approach to discover\,\nunderstand\, and interpret novel or not-well de fined patterns. It\nallows us to gain insights and generate hypotheses for subsequent\nanalyses. However\, visual exploration can become challenging when the\npatterns of interest are sparsely-distributed\, several orders of\nmagnitude smaller than the entire dataset\, or detected with high\nunc ertainty. In this talk\, I will discuss challenges in visually\nexploring multi-modal and multi-scale data\, and present new\nvisualization systems for efficiently browsing\, comparing\, and finding\npatterns in the contex t of genomic\, geospatial\, and time-series data.\nSpecifically\, I will d escribe a web platform for browsing multi-modal\nand multi-scale datasets\ , as well as their guided navigation. I will\npresent a generalized framew ork and toolkit for interactively\narranging\, grouping\, and aggregating thousands of pattern instances.\nAnd I will demonstrate how interactive vi sual machine learning can\nenhance our ability to find patterns effectivel y.\n\nhttps://datascience.utah.edu/talks/2021-02-12-fritz-lekschas/ URL:https://datascience.utah.edu/talks/2021-02-12-fritz-lekschas/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:biology & genomics,geospatial,machine learning,visualization END:VEVENT BEGIN:VEVENT UID:2021-02-19-samson-zhou@datascience.utah.edu DTSTAMP:20210219T210000Z DTSTART;TZID=America/Denver:20210219T140000 DTEND;TZID=America/Denver:20210219T150000 SUMMARY:Tight Bounds for Adversarially Robust Streams and Sliding Windows v ia Difference Estimators - Samson Zhou DESCRIPTION:Samson Zhou (CMU)\n\nWe introduce difference estimators for dat a stream computation\, which provide approximations to F(v)-F(u) for frequ ency vectors v\,u and a given function F. We show how to use such estimato rs to carefully trade error for memory in an iterative manner. The functio n F is generally non-linear\, and we give the first difference estimators for the frequency moments F_p for p between 0 and 2\, as well as for integ ers p>2. Using these\, we resolve a number of central open questions in ad versarial robust streaming and sliding window models.\n\nFor both models\, we obtain algorithms for norm estimation whose dependence on epsilon is 1 /epsilon^2\, which shows\, up to logarithmic factors\, that there is no ov erhead over the standard insertion-only data stream model for these proble ms.\n\nhttps://datascience.utah.edu/talks/2021-02-19-samson-zhou/ URL:https://datascience.utah.edu/talks/2021-02-19-samson-zhou/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE CATEGORIES:algorithms & theory,statistics END:VEVENT BEGIN:VEVENT UID:2021-02-26-rajesh-jayaram@datascience.utah.edu DTSTAMP:20210226T210000Z DTSTART;TZID=America/Denver:20210226T140000 DTEND;TZID=America/Denver:20210226T150000 SUMMARY:An Improved Analysis of the Quadtree for High Dimensional EMD - Raj esh Jayaram DESCRIPTION:Rajesh Jayaram (CMU)\n\nThe Earth Mover Distance (EMD) between two multi-sets A\,B in R^d of size s is the min-cost of bipartite matching s between points in A and B\, where cost is measured by distance between p oints. In this talk\, we discuss a classic divide-and-conquer algorithm kn own as Quadtree for approximating EMD.\nWe give a new analysis of the Quad tree\, showing that it gives a Õ(log s) approximation. This improves on t he previous known O(min{log s \, log d} * log s)-approximation of Andoni\, Indyk\, and Krauthgamer [SODA 08]\, and Backurs\, Dong\, Indyk\, Razensht eyn\, and Wagner (ICML 20).\n\nWe also give new space efficient sketching and streaming algorithms for estimating EMD with the improved approximatio n factor. The main conceptual contribution is an analytical framework for studying the Quadtree which goes beyond worst-case distortion of randomize d tree embeddings.\n\nBased on a joint work with Xi Chen\, Amit Levi\, and Erik Waingarten.\n\nPaper: https://rajeshjayaram.com/EarthMoverCJLW.pdf ( https://rajeshjayaram.com/EarthMoverCJLW.pdf)\n\nSlides: https://rajeshjay aram.com/EarthMoverCJLW.pdf\n\nhttps://datascience.utah.edu/talks/2021-02- 26-rajesh-jayaram/ URL:https://datascience.utah.edu/talks/2021-02-26-rajesh-jayaram/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:algorithms & theory,data management,machine learning END:VEVENT BEGIN:VEVENT UID:2021-03-12-yaoqing-yang@datascience.utah.edu DTSTAMP:20210312T210000Z DTSTART;TZID=America/Denver:20210312T140000 DTEND;TZID=America/Denver:20210312T150000 SUMMARY:Boundary thickness and robustness in learning models - Yaoqing Yang DESCRIPTION:Yaoqing Yang (UC Berkeley)\n\nRobustness of machine learning mo dels to various adversarial and non-adversarial corruptions continues to b e of interest. In this talk\, we present the notion of the "boundary thick ness" of a classifier\, and we describe its connection with and usefulness for model robustness. Thick decision boundaries lead to improved performa nce\, while thin decision boundaries lead to overfitting (e.g.\, measured by the robust generalization gap between training and testing) and lower r obustness. We show that a thicker boundary helps improve robustness agains t adversarial examples (e.g.\, improving the robust test accuracy of adver sarial training) as well as so-called out-of-distribution (OOD) transforms \, and we show that many commonly-used regularization and data augmentatio n procedures can increase boundary thickness. On the theoretical side\, we establish that maximizing boundary thickness during training is akin to t he so-called mixup training procedure. Using these observations\, we show that noise-augmentation on mixup training further increases boundary thick ness\, thereby combating vulnerability to various forms of adversarial att acks and OOD transforms. We can also show that the performance improvement in several lines of recent work happens in conjunction with a thicker bou ndary.\n\nhttps://datascience.utah.edu/talks/2021-03-12-yaoqing-yang/ URL:https://datascience.utah.edu/talks/2021-03-12-yaoqing-yang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:machine learning,privacy & security END:VEVENT BEGIN:VEVENT UID:2021-03-19-vivek-gupta@datascience.utah.edu DTSTAMP:20210319T200000Z DTSTART;TZID=America/Denver:20210319T140000 DTEND;TZID=America/Denver:20210319T150000 SUMMARY:Logic based classification for Low Resource Setting - Vivek Gupta DESCRIPTION:Vivek Gupta (University of Utah)\n\nAn NLP model’s ability to reason should be independent of language. Previous works utilize Natural Language Inference(NLI) to understand the reasoning ability of models\, mo stly focusing on high resource languages like English. To address scarcity of data in low-resource languages such as Hindi\, we use data recasting t o create four NLI datasets from existing four text classification datasets in Hindi language. Through experiments\, we show that our recasted datase t is devoid of statistical irregularities and spurious patterns. We study the consistency in predictions of the textual entailment models and propos e a consistency regulariser to remove pairwise-inconsistencies in predicti ons. Furthermore\, we propose a novel two-step classification method which uses textual-entailment predictions for classification tasks. We further improve the classification performance by jointly training the classificat ion and textual entailment tasks together. We therefore highlight the bene fits of data recasting and our approach with supporting experimental resul ts. You can access the dataset and paper here: https://www.aclweb.org/anth ology/2020.aacl-main.71.pdf (https://www.aclweb.org/anthology/2020.aacl-ma in.71.pdf). Joint work with BloomBerg AI.\n\nSlides: https://www.aclweb.or g/anthology/2020.aacl-main.71.pdf\n\nhttps://datascience.utah.edu/talks/20 21-03-19-vivek-gupta/ URL:https://datascience.utah.edu/talks/2021-03-19-vivek-gupta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVEtwWk5pemRu M0JsZz09 CATEGORIES:machine learning,natural language processing END:VEVENT BEGIN:VEVENT UID:2021-04-09-emily-beth-wall@datascience.utah.edu DTSTAMP:20210409T200000Z DTSTART;TZID=America/Denver:20210409T140000 DTEND;TZID=America/Denver:20210409T150000 SUMMARY:As We Are: Detecting and Mitigating Human Bias in Visual Analytics - Emily Beth Wall DESCRIPTION:Emily Beth Wall (Emory)\n\nVisual Analytics combines the comple mentary strengths of humans (perception and sensemaking capabilities) and machines (fast and accurate information processing). However\, people are susceptible to inherent limitations and biases\, including cognitive biase s (e.g.\, anchoring bias)\, social biases borne of cultural stereotypes an d prejudices (e.g.\, gender bias)\, and perceptual biases (e.g.\, illusion s). These biases can impact data analysis and decision making in critical ways\, leading to inaccurate or inefficient choices\, or even propagating long-standing institutional and systemic biases.\n\nGiven our knowledge of these biases and the increased use of data visualization to support decis ion making in data science\, the goal of this research is to detect and mi tigate human biases in visual data analysis. In this talk\, I describe (1) which types of bias are particularly relevant in the process of visual da ta analysis\, (2) how user interactions with data can be used to approxima te human biases\, and (3) how visualization systems can be designed to inc rease user awareness of potentially unconscious or implicit biases. By cre ating systems that promote real-time awareness of bias\, people can reflec t on their behavior and decision making and ultimately engage in a less-bi ased analysis and decision making process.\n\nhttps://datascience.utah.edu /talks/2021-04-09-emily-beth-wall/ URL:https://datascience.utah.edu/talks/2021-04-09-emily-beth-wall/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Virtual | https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVE twWk5pemRuM0JsZz09 CATEGORIES:fairness & ethics,visualization END:VEVENT BEGIN:VEVENT UID:2021-04-16-arun-sai-suggala@datascience.utah.edu DTSTAMP:20210416T200000Z DTSTART;TZID=America/Denver:20210416T140000 DTEND;TZID=America/Denver:20210416T150000 SUMMARY:Game Theoretic Statistics - Arun Sai Suggala DESCRIPTION:Arun Sai Suggala (CMU)\n\nGame theory and statistics are often regarded as disparate research areas. This is because typical statistical estimation settings are non-adversarial\, and the samples are assumed to b e generated by some stationary non-reactive source. However\, there is a g reat degree of commonality between the two fields. Classically\, the mathe matical philosophy of statistics\, particularly frequentist statistics\, p osits that the source of samples is potentially adversarial. This resulted in the rich theory of minimax statistical games and estimation. Boosting algorithms\, which are often regarded as best off-the-shelf classifiers\, can be viewed as playing a zero-sum game against a weak learner. To allow for various departures of ``test environment'' from ``train environments'' \, the emerging field of robust machine learning allows for adversarial ma nipulation of the train or test environments. Finally\, an emerging class of density estimators (GANs) in modern machine learning use an adversarial ``critic'' of the density estimator to improve the final density estimati on. The common theme among these classical and modern developments is an i nterplay between statistical estimation and two player games.\n\nIn this t alk\, I will present some of my recent work at the intersection of statist ics and game theory and show how game theory can help statistics. In parti cular\, I will present my work on minimax statistical estimation\, where w e develop algorithmic techniques for constructing minimax estimators. Our algorithms rely on tools from online nonconvex learning and help us constr uct minimax estimators for fundamental estimation problems such as covaria nce estimation and entropy estimation.\n\nhttps://datascience.utah.edu/tal ks/2021-04-16-arun-sai-suggala/ URL:https://datascience.utah.edu/talks/2021-04-16-arun-sai-suggala/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Virtual | https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVE twWk5pemRuM0JsZz09 CATEGORIES:algorithms & theory,machine learning,optimization,statistics END:VEVENT BEGIN:VEVENT UID:2021-04-23-lizzie-kumar@datascience.utah.edu DTSTAMP:20210423T200000Z DTSTART;TZID=America/Denver:20210423T140000 DTEND;TZID=America/Denver:20210423T150000 SUMMARY:Epistemic values in feature importance methods: Lessons from femini st epistemology - Lizzie Kumar DESCRIPTION:Lizzie Kumar (Utah)\n\nAs the public seeks greater accountabili ty and transparency from machine learning algorithms\, the research litera ture on methods to explain algorithms and their outputs has rapidly expand ed. Feature importance\, or the practice of assigning quantitative importa nce values to the input features of a machine learning model\, form a popu lar class of such methods. Much of the research on feature importance rest s on formalizations that attempt to capture universally desirable properti es. We investigate the ways in which epistemic values are implicitly embed ded in these methods and analyze the ways in which they conflict with idea s from feminist philosophy. We offer some suggestions on how to conduct re search on explanations that respects feminist epistemic values\, taking in to account the importance of social context\, the epistemic privileges of subjugated knowers\, and adopting more interactional ways of knowing.\n\nh ttps://datascience.utah.edu/talks/2021-04-23-lizzie-kumar/ URL:https://datascience.utah.edu/talks/2021-04-23-lizzie-kumar/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Virtual | https://us02web.zoom.us/j/87538638627?pwd=WGZrYmIyNjFsVE twWk5pemRuM0JsZz09 CATEGORIES:algorithms & theory,fairness & ethics,machine learning END:VEVENT BEGIN:VEVENT UID:2021-08-27-jeff-phillips@datascience.utah.edu DTSTAMP:20210827T200000Z DTSTART;TZID=America/Denver:20210827T140000 DTEND;TZID=America/Denver:20210827T150000 SUMMARY:A Visual tour of Bias Mitigation - Jeff Phillips DESCRIPTION:Jeff Phillips (University of Utah)\n\nWord vector embeddings ha ve been shown to contain and amplify biases in data they are extracted fro m. Consequently\, many techniques have been proposed to identify\, mitigat e\, and attenuate these biases in word representations. In this talk\, I w ill review a collection of state-of-the-art debiasing techniques. To aid t his\, we provide an open source web-based visualization tool VERB (Visuali zation of Embedding Representations for deBiasing) and offer hands-on expe rience in exploring the effects of these debiasing techniques on the geome try of high-dimensional word vectors. To help understand how various debia sing techniques change the underlying geometry\, I will show how to decomp ose each technique into interpretable sequences of primitive operations an d study their effect on the word vectors using dimensionality reduction an d interactive visual exploration.\n\nhttps://datascience.utah.edu/talks/20 21-08-27-jeff-phillips/ URL:https://datascience.utah.edu/talks/2021-08-27-jeff-phillips/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hG V1hTK05SUFZwUT09 CATEGORIES:deep learning,fairness & ethics,natural language processing,visu alization END:VEVENT BEGIN:VEVENT UID:2021-09-03-shandian-zhe@datascience.utah.edu DTSTAMP:20210903T200000Z DTSTART;TZID=America/Denver:20210903T140000 DTEND;TZID=America/Denver:20210903T150000 SUMMARY:Multi-fidelity Learning and Optimization for Physical Simulation an d AutoML - Shandian Zhe DESCRIPTION:Shandian Zhe (Utah SoC)\n\nMulti-fidelity learning involves usi ng training examples at different fidelities or resolutions. High-fidelity examples are of high-quality but often are much more costly to collect th an inaccurate\, low-fidelity examples. How to retrieve and leverage exampl es at multiple fidelities is the key to reduce the learning cost while max imizing efficiency.\n\nThis talk will introduce our recent work in multi-f idelity learning and optimization. First\, I will introduce our deep auto- regressive models that can capture complex correlations across the fidelit ies to integrate examples of high-dimensional outputs. These are common in applications of physical simulation. Second\, I will introduce our work o f deep multi-fidelity active learning and Bayesian optimization that can i mprove the learning and optimization efficiency while reducing the cost of generating training examples\, namely maximizing the benefit-cost ratio. Finally\, a batch version of the active learning and optimization techniqu e will be presented\, which can reduce the query redundancy\, improve dive rsity\, and further boost the benefit-cost ratio. I will showcase the adva ntage of our methods in standard benchmarks of physical simulation\, topol ogy structure optimization\, and typical tasks in hyper-parameter tuning/A utoML.\n\nhttps://datascience.utah.edu/talks/2021-09-03-shandian-zhe/ URL:https://datascience.utah.edu/talks/2021-09-03-shandian-zhe/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:optimization END:VEVENT BEGIN:VEVENT UID:2021-09-10-pierre-lermusiaux@datascience.utah.edu DTSTAMP:20210910T200000Z DTSTART;TZID=America/Denver:20210910T140000 DTEND;TZID=America/Denver:20210910T150000 SUMMARY:Neural Closure Models for Dynamical Systems - Pierre Lermusiaux DESCRIPTION:Pierre Lermusiaux\n\nComplex dynamical systems are used for pre dictions in many domains. Because of computational costs\, models are trun cated\, coarsened or aggregated. As the neglected and unresolved terms bec ome important\, the utility of model predictions diminishes. We develop a novel\, versatile and rigorous methodology to learn non-Markovian closure parametrizations for known-physics/low-fidelity models using data from hig h-fidelity simulations. The new neural closure models augment low-fidelity models with neural delay differential equations (nDDEs)\, motivated by th e Mori–Zwanzig formulation and the inherent delays in complex dynamical systems. We demonstrate that neural closures efficiently account for trunc ated modes in reduced-order models\, capture the effects of subgrid-scale processes in coarse models and augment the simplification of complex biolo gical and physical-biogeochemical models. We find that using non-Markovian over Markovian closures improves long-term prediction accuracy and requir es smaller networks. We derive adjoint equations and network architectures needed to efficiently implement the new discrete and distributed nDDEs\, for any time-integration schemes and allowing non-uniformly spaced tempora l training data. The performance of discrete over-distributed delays in cl osure models is explained using information theory\, and we find an optima l amount of past information for a specified architecture. Finally\, we an alyze computational complexity and explain the limited additional cost due to neural closure models.\n\nPaper: https://royalsocietypublishing.org/do i/10.1098/rspa.2020.1004 (https://royalsocietypublishing.org/doi/10.1098/r spa.2020.1004?fbclid=IwAR1DZy5lHuf68lYwkk5Z_kUYzbHXdgRi0zUhoy5bohGy-w0BjZu wjqPhytQ)\n\nhttps://datascience.utah.edu/talks/2021-09-10-pierre-lermusia ux/ URL:https://datascience.utah.edu/talks/2021-09-10-pierre-lermusiaux/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hGV1hTK05SUFZ wUT09 CATEGORIES:algorithms & theory,machine learning,statistics END:VEVENT BEGIN:VEVENT UID:2021-09-17-ross-whitaker@datascience.utah.edu DTSTAMP:20210917T200000Z DTSTART;TZID=America/Denver:20210917T140000 DTEND;TZID=America/Denver:20210917T150000 SUMMARY:Air Quality Mapping Using Sensor Networks and Statistical Regressio n - Ross Whitaker DESCRIPTION:Ross Whitaker (Utah SoC & SCI)\n\nThis talk describes an air qu ality mapping system that is currently deployed in the Salt Lake Valley. W e begin with the motivations for the system and then briefly describe the cyberphysical infrastructure. We then review the Gaussian process modeling approach we are using and discuss several important practical considerati ons around challenges of data wrangling and numerical implementations. We present some examples of AQ estimates associated with specific events of b ad air quality. Finally we talk about current directions in research assoc iated with this approach.\n\nCOI Disclaimer: Ross Whitaker has a financial interest in the company Tetrad\, which has business interests related to the technologies in this talk.\n\nMEB 3147\n\nhttps://datascience.utah.edu /talks/2021-09-17-ross-whitaker/ URL:https://datascience.utah.edu/talks/2021-09-17-ross-whitaker/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 CATEGORIES:climate & environment,geospatial,statistics END:VEVENT BEGIN:VEVENT UID:2021-09-24-c-seshadhri@datascience.utah.edu DTSTAMP:20210924T200000Z DTSTART;TZID=America/Denver:20210924T140000 DTEND;TZID=America/Denver:20210924T150000 SUMMARY:Studying the (in)effectiveness of low dimensional graph embeddings - C. Seshadhri DESCRIPTION:C. Seshadhri (UC Santa Cruz)\n\nLow dimensional graph embedding s are a fundamental and popular tool used for machine learning on graphs. Given a graph\, the basic idea is to produce a low-dimensional vector for each vertex\, such that "similarity" in geometric space corresponds to "pr oximity" in the graph. These vectors can then be used as features in a ple thora of machine learning tasks\, such as link prediction\, community labe ling\, recommendations\, etc. Despite many results emerging in this area o ver the past few years\, there is less study on the core premise of these embeddings. Can such low-dimensional embeddings effectively capture the st ructure of real-world (such as social) networks? Contrary to common wisdom \, we mathematically prove and empirically demonstrate that popular low-di mensional graph embeddings do not capture salient properties of real-world networks. We mathematically prove that common low-dimensional embeddings cannot generate graphs with both low average degree and large clustering c oefficients\, which have been widely established to be empirically true fo r real-world networks. Empirically\, we observe that the embeddings genera ted by popular methods fail to recreate the triangle structure of real-wor ld networks\, and do not perform well on certain community labeling tasks. \n\nJoint work with Ashish Goel\, Caleb Levy\, Aneesh Sharma\, and Andrew Stolman\n\nhttps://datascience.utah.edu/talks/2021-09-24-c-seshadhri/ URL:https://datascience.utah.edu/talks/2021-09-24-c-seshadhri/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hGV1hTK05SUFZ wUT09 CATEGORIES:algorithms & theory,deep learning,machine learning,networks & gr aphs END:VEVENT BEGIN:VEVENT UID:2021-10-01-tony-h-grubesic@datascience.utah.edu DTSTAMP:20211001T200000Z DTSTART;TZID=America/Denver:20211001T140000 DTEND;TZID=America/Denver:20211001T150000 SUMMARY:Estimating Potential Oil Spill Trajectories and Coastal Impacts fro m Near-Shore Storage Facilities: A Case Study of FSO Nabarima and the Gulf of Paria - Tony H. Grubesic DESCRIPTION:Tony H. Grubesic (UT Austin)\n\nThe FSO Nabarima is a floating storage facility and offloading vessel in the Gulf of Paria\, between Vene zuela and the island of Trinidad. During the latter half of 2020\, the Nab arima was disabled\, holding approximately 1.3 million barrels (55 million gallons) of crude oil on board. In October of 2020\, the vessel was tilti ng and potentially at risk of spilling its payload into open water. Althou gh all of the oil on the Nabarima was successfully offloaded by April 2021 \, the threat of large crude oil releases is ubiquitous and persistent in many coastal regions\, threatening local ecosystems and livelihoods in coa stal communities. The purpose of this presentation is to highlight a geoco mputational framework for evaluating the potential spatial vulnerability o f coastlines should oil be released from near-shore storage facilities. We use the Nabarima as a broadly representative case study\, discuss potenti al spill cleanup and mitigation strategies\, and highlight the challenges of coordinating cross-national responses to these types of spill scenarios .\n\nhttps://datascience.utah.edu/talks/2021-10-01-tony-h-grubesic/ URL:https://datascience.utah.edu/talks/2021-10-01-tony-h-grubesic/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hGV1hTK05SUFZ wUT09 CATEGORIES:society & policy END:VEVENT BEGIN:VEVENT UID:2021-10-08-anna-little@datascience.utah.edu DTSTAMP:20211008T200000Z DTSTART;TZID=America/Denver:20211008T140000 DTEND;TZID=America/Denver:20211008T150000 SUMMARY:The Mathematics of the Signal-to-Noise Ratio and Insights for Data Science - Anna Little DESCRIPTION:Anna Little (Utah Math)\n\nDespite the huge variety of data typ es and goals in data science\, there is usually an underlying signal-to-no ise ratio that governs the difficulty of the data science task. Analysis o f this signal-to-noise ratio leads to important insights about sample size requirements and the impact of the data dimension\, which are consistent across various data models and tasks. This talk will illustrate these univ ersal insights in three specific contexts: (1) density-based clustering vi a graph embeddings\, (2) clustering mixture models via multidimensional sc aling\, and (3) signal recovery from noisy data. The underlying data model s are motivated by applications such as imaging\, particle physics\, singl e-cell RNA sequencing\, and cryo-electron microscopy.\n\nhttps://datascien ce.utah.edu/talks/2021-10-08-anna-little/ URL:https://datascience.utah.edu/talks/2021-10-08-anna-little/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hG V1hTK05SUFZwUT09 CATEGORIES:algorithms & theory,biology & genomics END:VEVENT BEGIN:VEVENT UID:2021-10-22-erin-wolf-chambers@datascience.utah.edu DTSTAMP:20211022T200000Z DTSTART;TZID=America/Denver:20211022T140000 DTEND;TZID=America/Denver:20211022T150000 SUMMARY:Applications of topology and geometry to root analysis - Erin Wolf Chambers DESCRIPTION:Erin Wolf Chambers (St. Loius University)\n\nAnalysis of 3d sha pes is a core problem in many fields\, and there are\nmany tools from topo logy and geometry that can provide insight and\nunderstanding. In this tal k\, we focus on developing significance\nmeasures for 3d plant structures\ , primarily root systems of plants.\nOur measures are based on the medial axis transform\, which plays a\nfundamental role in shape matching and ana lysis\, but is widely known\nto be unstable to even small boundary perturb ations. Methods for\npruning the medial axis are usually guided by some me asure of\nsignificance\, with considerable work done for both 2- and\n3-di mensional shapes. Such significance measures can be used for\nidentifying salient features\, and hence are useful for simplification\,\ncomparison\, and alignment. In this talk\, we will present theoretical\ninsights and p roperties of commonly used significance measures\,\nfocusing on those in 2 D and 3D that are both shape-revealing and\ntopology-preserving\, as well as being robust to noise on the boundary.\nWe'll then discuss several meth ods that de-noise a shape and identify\ntopologically and geometrically pr ominent features\, using both the\nmedial axis and other measures commonly used in topological data\nanalysis. Our methods are quite successful comp ared to the state of\nthe art\, and are available in the package TopoRoot\ , an automatic\npipeline for plant architectural analysis from 3D Imaging. \n\nhttps://datascience.utah.edu/talks/2021-10-22-erin-wolf-chambers/ URL:https://datascience.utah.edu/talks/2021-10-22-erin-wolf-chambers/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hG V1hTK05SUFZwUT09 END:VEVENT BEGIN:VEVENT UID:2021-10-29-bao-wang@datascience.utah.edu DTSTAMP:20211029T200000Z DTSTART;TZID=America/Denver:20211029T140000 DTEND;TZID=America/Denver:20211029T150000 SUMMARY:How Differential Equations and Random Graph Insights Benefit Deep L earning - Bao Wang DESCRIPTION:Bao Wang (Utah Math\, SCI)\n\nWe will present recent results on developing new deep learning algorithms leveraging differential equations and random graph insights.\nFirst\, we will present a new class of contin uous-depth deep neural networks that were motivated by the ODE limit of th e classical momentum method\, named heavy-ball neural ODEs (HBNODEs). HBNO DEs enjoy two properties that imply practical advantages over NODEs: (i) T he adjoint state of an HBNODE also satisfies an HBNODE\, accelerating both forward and backward ODE solvers\, thus significantly accelerate learning and improve the utility of the trained models. (ii) The spectrum of HBNOD Es is well structured\, enabling effective learning of long-term dependenc ies from complex sequential data.\nSecond\, we will extend HBNODE to graph learning leveraging diffusion on graphs\, resulting in new algorithms for deep graph learning. The new algorithms are more accurate than existing d eep graph learning algorithms and more scalable to deep architectures\, an d also suitable for learning at low labeling rate regimes. Moreover\, we w ill present a fast multipole method-based efficient attention mechanism fo r modeling graph nodes interactions.\nThird\, if time permits\, we will di scuss building an efficient and reliable overlay network for decentralized federated learning based on the random graph theory.\n\nhttps://datascien ce.utah.edu/talks/2021-10-29-bao-wang/ URL:https://datascience.utah.edu/talks/2021-10-29-bao-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hG V1hTK05SUFZwUT09 CATEGORIES:algorithms & theory,deep learning,networks & graphs END:VEVENT BEGIN:VEVENT UID:2021-11-05-marina-kogan@datascience.utah.edu DTSTAMP:20211105T200000Z DTSTART;TZID=America/Denver:20211105T140000 DTEND;TZID=America/Denver:20211105T150000 SUMMARY:Sequence-based approaches as human-centered data science methods fo r crisis informatics - Marina Kogan DESCRIPTION:Marina Kogan (Utah SoC)\n\nSocial media platforms have been inc reasingly used by the public in crisis situations\, partly because they up end the traditional top-down broadcasting model of risk communication. Ins tead\, social media platforms facilitate a two-way information exchange be tween the official response channels and the general public\, enabling mor e participatory crisis communication\, as well as coordination and self-or ganization among the public. In this more complex information ecosystem\, understanding the flow of information is crucial to supporting those affec ted and preventing malicious actors from capitalizing on the uncertainty. However\, the study of such information flows is challenging\, because the high-tempo\, high-volume convergent nature of crisis events produces vast amounts of social media data\, necessitating the use of the data science methods. On the other hand\, to glean meaningful insight from the crisis-r elated social media activity\, it is necessary to use methods that account for the complex social context of the user activity. In this talk I will show how the Human-Centered Data Science (HCDS) provides methodological ap proaches that both harness the power of computational methods and account for the highly situated nature of social media activity in disruption. I w ill focus on sequence-based approaches as examples of HCDS methods in two empirical studies: analysis of attention-garnering information during a na tural disaster and investigation of behavioral signatures in coordinated i nformation operations.\n\nhttps://datascience.utah.edu/talks/2021-11-05-ma rina-kogan/ URL:https://datascience.utah.edu/talks/2021-11-05-marina-kogan/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hG V1hTK05SUFZwUT09 CATEGORIES:human-centered computing,society & policy END:VEVENT BEGIN:VEVENT UID:2021-11-12-mikhail-belkin@datascience.utah.edu DTSTAMP:20211112T210000Z DTSTART;TZID=America/Denver:20211112T140000 DTEND;TZID=America/Denver:20211112T150000 SUMMARY:From classical statistics to modern deep learning - Mikhail Belkin DESCRIPTION:Mikhail Belkin\n\nRecent empirical successes of deep learning h ave exposed significant gaps in our\nfundamental understanding of learning and optimization mechanisms.\nModern best practices for model selection a re in direct contradiction to the methodologies\nsuggested by classical an alyses. Similarly\, the efficiency of SGD-based local methods\nused in tra ining modern models\, appeared at odds with the standard intuitions on opt imization.\n\nFirst\, I will present evidence\, empirical and mathematical \, that necessitates\nrevisiting classical statistical notions\, such as o ver-fitting. I will continue to discuss the emerging\nunderstanding of gen eralization\, and\, in particular\, the "double descent" risk curve\, whic h extends\nthe classical U-shaped generalization curve beyond the point of interpolation.\n\nSecond\, I will discuss why the landscapes of over-para meterized neural networks are\ngenerically never convex\, even locally. In stead they satisfy the Polyak-Lojasiewicz (PL)\ncondition across most of t he parameter space instead\, presents an powerful framework for optimizati on in general over-parameterized models and allows SGD-type methods to con verge to a global minimum.\n\nWhile our understanding has significantly gr own in the last few years\, a key piece of the puzzle remains -- how does optimization align with statistics to form the complete mathematical pictu re of modern ML?\n\nhttps://datascience.utah.edu/talks/2021-11-12-mikhail- belkin/ URL:https://datascience.utah.edu/talks/2021-11-12-mikhail-belkin/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hGV1hTK05SUFZ wUT09 CATEGORIES:deep learning,machine learning,optimization,statistics END:VEVENT BEGIN:VEVENT UID:2021-11-19-michael-yeh@datascience.utah.edu DTSTAMP:20211119T210000Z DTSTART;TZID=America/Denver:20211119T140000 DTEND;TZID=America/Denver:20211119T150000 SUMMARY:Towards a Near Universal Time Series Data Mining Tool: Introducing the Matrix Profile - Michael Yeh DESCRIPTION:Michael Yeh (Visa Research)\n\nMatrix profile is a data structu re that annotates a time series by recording the location of and the dista nce to the nearest neighbors of each subsequences in the time series. The matrix profile stores such information in an efficient and easy-to-access fashion and can be used in a variety of data mining tasks like motif/disco rd discovery\, semantic segmentation\, and clustering. In this talk\, I wi ll 1) introduce what matrix profile is\, 2) discuss the computational chal lenge associated with matrix profile\, and 3) show how it can be used in d ifferent time series data mining tasks.\n\nhttps://datascience.utah.edu/ta lks/2021-11-19-michael-yeh/ URL:https://datascience.utah.edu/talks/2021-11-19-michael-yeh/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hGV1hTK05SUFZ wUT09 CATEGORIES:algorithms & theory,computer vision END:VEVENT BEGIN:VEVENT UID:2021-12-03-julia-silge@datascience.utah.edu DTSTAMP:20211203T210000Z DTSTART;TZID=America/Denver:20211203T140000 DTEND;TZID=America/Denver:20211203T150000 SUMMARY:Data visualization for machine learning practitioners - Julia Silge DESCRIPTION:Julia Silge (RStudio)\n\nVisual representations of data inform how machine learning practitioners think\, understand\, and decide. Before charts are ever used for outward communication about a ML system\, they a re used by the system designers and operators themselves as a tool to make better modeling choices. Practitioners use visualization\, from very fami liar statistical graphics to creative and less standard plots\, at the poi nts of most important human decisions when other ways to validate those de cisions can be difficult. Visualization approaches are used to understand both the data that serves as input for machine learning and the models tha t practitioners create. In this talk\, learn about the process of building a ML model in the real world\, how and when practitioners use visualizati on to make more effective choices\, and considerations for ML visualizatio n tooling.\n\nhttps://datascience.utah.edu/talks/2021-12-03-julia-silge/ URL:https://datascience.utah.edu/talks/2021-12-03-julia-silge/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 and Zoom | https://utah.zoom.us/j/93778940103?pwd=TStQRWh WVjRxd0hGV1hTK05SUFZwUT09 CATEGORIES:human-centered computing,machine learning,visualization END:VEVENT BEGIN:VEVENT UID:2021-12-10-sameer-singh@datascience.utah.edu DTSTAMP:20211210T210000Z DTSTART;TZID=America/Denver:20211210T140000 DTEND;TZID=America/Denver:20211210T150000 SUMMARY:Evaluating and Testing Natural Language Processing Models - Sameer Singh DESCRIPTION:Sameer Singh (UC Irvine)\n\nCurrent evaluation of the generaliz ation of natural language processing (NLP) systems\, and much of machine l earning\, primarily consists of measuring the accuracy on held-out instanc es of the dataset. Since the held-out instances are often gathered using s imilar annotation process as the training data\, they include the same bia ses that act as shortcuts for machine learning models\, allowing them to a chieve accurate results without requiring actual natural language understa nding. Thus held-out accuracy is often a poor proxy for measuring generali zation. Further\, aggregate metrics have little to say about where the pro blems may lie\, and how to address them.\nIn this talk\, I will introduce a number of approaches we are investigating to perform a more thorough eva luation of NLP systems. I will first provide a quick overview of automated techniques for perturbing instances in the dataset that identify loophole s and shortcuts in NLP models\, including semantic adversaries and univers al triggers. I will then describe recent work on creating comprehensive an d thorough tests and evaluation benchmarks for NLP using CheckList\, that aim to directly evaluate comprehension and understanding capabilities. The talk will include a number of NLP tasks\, such as sentiment analysis\, te xtual entailment\, paraphrase detection\, and question answering.\n\nDr. S ameer Singh is an Associate Professor of Computer Science at the Universit y of California\, Irvine (UCI) and an Allen AI Fellow at Allen Institute f or AI. He is working primarily on robustness and interpretability of machi ne learning algorithms\, along with models that reason with text and struc ture for natural language processing. Sameer was a postdoctoral researcher at the University of Washington and received his PhD from the University of Massachusetts\, Amherst. He has received the NSF CAREER award\, selecte d as a DARPA Riser\, UCI Distinguished Early Career Faculty award\, and th e Hellman Faculty Fellowship. His group has received funding from Allen In stitute for AI\, Amazon\, NSF\, DARPA\, Adobe Research\, Hasso Plattner In stitute\, NEC\, Base 11\, and FICO. Sameer has published extensively at ma chine learning and natural language processing venues and received confere nce paper awards at KDD 2016\, ACL 2018\, EMNLP 2019\, AKBC 2020\, and ACL 2020. (https://sameersingh.org/)\n\nhttps://datascience.utah.edu/talks/20 21-12-10-sameer-singh/ URL:https://datascience.utah.edu/talks/2021-12-10-sameer-singh/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/93778940103?pwd=TStQRWhWVjRxd0hGV1hTK05SUFZ wUT09 CATEGORIES:algorithms & theory,machine learning,natural language processing END:VEVENT BEGIN:VEVENT UID:2022-01-14-yi-zhou@datascience.utah.edu DTSTAMP:20220114T220000Z DTSTART;TZID=America/Denver:20220114T150000 DTEND;TZID=America/Denver:20220114T160000 SUMMARY:Understanding the Convergence of Optimization Algorithms for Minima x Machine Learning - Yi Zhou DESCRIPTION:Yi Zhou (Utah ECE)\n\nThe past decade has witnessed the great s uccess of deep learning in broad societal and commercial applications. How ever\, conventional deep learning relies on wildly fitting data with neura l networks\, which is known to produce models that lack resilience. For in stance\, models used in facial recognition and healthcare are known to be biased toward people of a certain race or gender. Models used in autonomou s driving are vulnerable to malicious attacks\, i.e.\, putting an art stic ker on a stop sign may force the model to classify it as a speed limit sig n. Therefore\, the next-generation deep learning paradigm aims to deliver resilient models that promote robustness to malicious attacks\, fairness a mong users\, and privacy preservation\, and this can be realized by levera ging the emerging minimax machine learning framework. In this talk\, I wil l present three gradient-descent-ascent (GDA) type of optimization algorit hms for solving different classes of nonconvex minimax machine learning pr oblems. Then\, I will present a principled nonconvex minimax optimization theory that establishes the global convergence and convergence rates of th ese algorithms.\n\nhttps://datascience.utah.edu/talks/2022-01-14-yi-zhou/ URL:https://datascience.utah.edu/talks/2022-01-14-yi-zhou/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1250 | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeFpQN0lB MEZQQ2Z1ZUdFUT09 CATEGORIES:algorithms & theory,deep learning,machine learning,optimization END:VEVENT BEGIN:VEVENT UID:2022-01-21-chinmay-hedge@datascience.utah.edu DTSTAMP:20220121T220000Z DTSTART;TZID=America/Denver:20220121T150000 DTEND;TZID=America/Denver:20220121T160000 SUMMARY:Designing Neural Networks for Efficient Encrypted Inference - Chinm ay Hedge DESCRIPTION:Chinmay Hedge (NYU)\n\nAs deep neural networks become ever more pervasive\, so too are concerns surrounding users' data privacy. Curiousl y\, standard cryptographic encryption approaches for guaranteeing data pri vacy do not interact well with traditional neural network models. In this talk\, I will (a) outline why standard networks are not encryption-efficie nt\, (b) suggest two new approaches for designing deep networks that do su pport efficient and secure inference\, and (c) show results instantiating these approaches on real-world use cases.\n\nhttps://datascience.utah.edu/ talks/2022-01-21-chinmay-hedge/ URL:https://datascience.utah.edu/talks/2022-01-21-chinmay-hedge/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeFpQN0lBMEZQQ2Z1ZUd FUT09 CATEGORIES:deep learning,privacy & security END:VEVENT BEGIN:VEVENT UID:2022-01-28-swaroop-mishra@datascience.utah.edu DTSTAMP:20220128T220000Z DTSTART;TZID=America/Denver:20220128T150000 DTEND;TZID=America/Denver:20220128T160000 SUMMARY:Towards the Development of Models that Learn New Tasks from Instruc tions - Swaroop Mishra DESCRIPTION:Swaroop Mishra (ASU\, Ph.D. Student)\n\nhttps://datascience.uta h.edu/talks/2022-01-28-swaroop-mishra/ URL:https://datascience.utah.edu/talks/2022-01-28-swaroop-mishra/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:education END:VEVENT BEGIN:VEVENT UID:2022-02-04-tuhin-chakrabarty@datascience.utah.edu DTSTAMP:20220204T220000Z DTSTART;TZID=America/Denver:20220204T150000 DTEND;TZID=America/Denver:20220204T160000 SUMMARY:The Curious Case of Figurative Language - Tuhin Chakrabarty DESCRIPTION:Tuhin Chakrabarty (Columbia University\, Ph.D Student)\n\nDespi te the ubiquity of figurative language across various forms of speech and writing\, the vast majority of NLP research focuses primarily on literal l anguage. Figurative language is challenging because of its implicit nature . In this talk\, I will specifically focus on the following questions: 1) Can large language models understand/interpret them? 2) Can computers gene rate figurative language?\n\nhttps://datascience.utah.edu/talks/2022-02-04 -tuhin-chakrabarty/ URL:https://datascience.utah.edu/talks/2022-02-04-tuhin-chakrabarty/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:natural language processing END:VEVENT BEGIN:VEVENT UID:2022-02-11-vaggos-chatziafratis@datascience.utah.edu DTSTAMP:20220211T220000Z DTSTART;TZID=America/Denver:20220211T150000 DTEND;TZID=America/Denver:20220211T160000 SUMMARY:Neural Networks Expressivity through the lens of Dynamical Systems - Vaggos Chatziafratis DESCRIPTION:Vaggos Chatziafratis (UC Santa Cruz)\n\nGiven a target function f\, how large must a neural network be in order to approximate f?\nUnders tanding the representational power of Deep Neural Networks (DNNs) and how their structural properties (e.g.\, depth\, width\, type of activation uni t) affect the functions they can compute\, has been an important yet chall enging question in approximation theory and deep learning even in the earl y days of AI.\n\nIn this talk\, I want to tell you about some recent progr ess on this topic that uses ideas from dynamical systems. The main results are exponential depth-width trade-offs for DNNs representing certain fami lies of functions. Our techniques rely on a generalized notion of fixed po ints\, called periodic points that have played a major role in chaos theor y (Li-Yorke chaos and Sharkovsky's theorem).\n\nBased on three recent work s:\n- with Ioannis Panageas\, Sai Ganesh Nagarajan and Xiao Wang from ICLR '20 (spotlight): https://arxiv.org/abs/1912.04378 (https://arxiv.org/abs/1 912.04378)\n- with Ioannis Panageas and Sai Ganesh Nagarajan from ICML'20: https://arxiv.org/abs/2003.00777 (https://arxiv.org/abs/2003.00777)\n- wi th Clayton Sanford from AISTATS'22: https://arxiv.org/abs/2110.10295 (http s://arxiv.org/abs/2110.10295)\n\nhttps://datascience.utah.edu/talks/2022-0 2-11-vaggos-chatziafratis/ URL:https://datascience.utah.edu/talks/2022-02-11-vaggos-chatziafratis/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeFpQN0lBMEZQQ2Z1ZUd FUT09 CATEGORIES:deep learning END:VEVENT BEGIN:VEVENT UID:2022-02-18-jeff-phillips@datascience.utah.edu DTSTAMP:20220218T220000Z DTSTART;TZID=America/Denver:20220218T150000 DTEND;TZID=America/Denver:20220218T160000 SUMMARY:Some Very Basic Theory of Classification (in low dimensions - Jeff Phillips DESCRIPTION:Jeff Phillips (Utah SoC)\n\nI plan to talk about some new resul ts on some very fundamental (but overlooked) theory questions in classific ation. First\, how fast can you find a linear classifier that eps-approxim ates the optimal one in terms of miss-classification? Second\, if you want to preserve a Euclidean margin between perfectly classified points for a polynomial classifier\, how many samples do you need?\n\nhttps://datascien ce.utah.edu/talks/2022-02-18-jeff-phillips/ URL:https://datascience.utah.edu/talks/2022-02-18-jeff-phillips/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1250 CATEGORIES:machine learning END:VEVENT BEGIN:VEVENT UID:2022-02-25-sunipa-dev@datascience.utah.edu DTSTAMP:20220225T220000Z DTSTART;TZID=America/Denver:20220225T150000 DTEND;TZID=America/Denver:20220225T160000 SUMMARY:Towards Inclusive and Socially Aware Language Technologies - Sunipa Dev DESCRIPTION:Sunipa Dev (Google Research)\n\nLarge language models are commo nly used in different paradigms of natural language processing and machine learning\, and are known for their efficiency as well as their overall la ck of interpretability. Their data driven approach for emulating human lan guage often results in human biases being encoded and even amplified\, pot entially leading to cyclic propagation of representational and allocationa l harm. We discuss in this talk some aspects of detecting\, evaluating\, a nd mitigating biases and associated harms in a holistic\, inclusive\, and culturally-aware manner. In particular\, we discuss the disparate impact o n society of common language tools that are not inclusive of all gender id entities.\n\nhttps://datascience.utah.edu/talks/2022-02-25-sunipa-dev/ URL:https://datascience.utah.edu/talks/2022-02-25-sunipa-dev/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeFpQN0lBMEZQQ2Z1ZUd FUT09 CATEGORIES:fairness & ethics,natural language processing END:VEVENT BEGIN:VEVENT UID:2022-03-04-anirudh-goyal-of-montreal@datascience.utah.edu DTSTAMP:20220304T220000Z DTSTART;TZID=America/Denver:20220304T150000 DTEND;TZID=America/Denver:20220304T160000 SUMMARY:From Specialists to Generalists: Inductive Biases of Deep Learning for Higher Level Cognition - Anirudh Goyal of Montreal DESCRIPTION:Anirudh Goyal of Montreal\n\nA fascinating hypothesis is that h uman and animal intelligence could be explained by a few principles (rathe r than an encyclopedic list of heuristics). If that hypothesis was correct \, we could more easily both understand our own intelligence and build int elligent machines. Just like in physics\, the principles themselves would not be sufficient to predict the behavior of complex systems like brains\, and substantial computation might be needed to simulate human-like intell igence. This hypothesis would suggest that studying the kind of inductive biases that humans and animals exploit could help both clarify these princ iples and provide inspiration for AI research and neuroscience theories. D eep learning already exploits several key inductive biases\, and my work c onsiders a larger list\, focusing on those which concern mostly higher-lev el and sequential conscious processing. The objective of clarifying these particular principles is that they could potentially help us build AI syst ems benefiting from humans' abilities in terms of flexible out-of-distribu tion and systematic generalization\, which is currently an area where a la rge gap exists between state-of-the-art machine learning and human intelli gence.\n\nhttps://datascience.utah.edu/talks/2022-03-04-anirudh-goyal-of-m ontreal/ URL:https://datascience.utah.edu/talks/2022-03-04-anirudh-goyal-of-montreal / SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:deep learning,fairness & ethics,machine learning,statistics END:VEVENT BEGIN:VEVENT UID:2022-03-18-greg-herschlag-jon-mattingly@datascience.utah.edu DTSTAMP:20220318T210000Z DTSTART;TZID=America/Denver:20220318T150000 DTEND;TZID=America/Denver:20220318T160000 SUMMARY:Quantifying Gerrymandering: Hearing the Will of the People - Greg H erschlag\, Jon Mattingly DESCRIPTION:Greg Herschlag (Duke University)\, Jon Mattingly (Duke Universi ty)\n\nGerrymandering is the process of manipulating political\ndistricts either to amplify the power of a political group or suppress\nthe represen tation of certain demographic groups. Although we have seen\nincreasingly precise and effective gerrymanders\, a number of\nmathematicians\, politic al scientists\, and lawyers are developing\neffective methodologies at unc overing and understanding the intent and\neffects of gerrymandered distric ts.\n\nThe basic idea behind these methods is to compare a given set of\nd istricts to a large collection of neutrally drawn plans. The process\nreli es on three distinct components: First\, we determine rules for\ncompliant redistricting plans along with codifying preferences between\nthese plans \; next\, we sample the space of compliant redistricting plans\n(according to our preferences) and generate a large collection of\nnon-partisan alte rnatives\; finally\, we compare the collection of plans\nto a particular p lan of interest.\n\nIn this talk\, we will discuss how our research group at Duke has\nanalyzed gerrymandering. We will discuss the sampling methods we employ\,\nincluding several recent algorithmic advances. We will also discuss how\nrecent advances in North Carolina's legal theory have brought partisan\nsymmetry considerations back into focus.\n\nhttps://datascience .utah.edu/talks/2022-03-18-greg-herschlag-jon-mattingly/ URL:https://datascience.utah.edu/talks/2022-03-18-greg-herschlag-jon-mattin gly/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:algorithms & theory,statistics END:VEVENT BEGIN:VEVENT UID:2022-03-25-chad-topaz-jude-higdon@datascience.utah.edu DTSTAMP:20220325T210000Z DTSTART;TZID=America/Denver:20220325T150000 DTEND;TZID=America/Denver:20220325T160000 SUMMARY:Quantitative Approaches to Social Justice - Chad Topaz\, Jude Higdo n DESCRIPTION:Chad Topaz (QSIDE)\, Jude Higdon (QSIDE)\n\nCivil rights leader \, educator\, and investigative journalist Ida B. Wells said that "the way to right wrongs is to shine the light of truth upon them." This talk will demonstrate how mathematical\, statistical\, and computational approaches can shine a light on social injustices and help build solutions to remedy them. We will present research-to-action projects on diversity in art mus eums\, inclusion in STEM\, equity in criminal sentencing\, and other topic s. The tools engaged include crowdsourcing\, data cleaning\, clustering\, hypothesis testing\, statistical modeling\, Markov chains\, data visualiza tion\, and much more. Overall\, we hope that this talk leaves you informed about the breadth of social justice applications that one can tackle usin g quantitative tools in careful collaboration with other scholars and acti vists.\n\nhttps://datascience.utah.edu/talks/2022-03-25-chad-topaz-jude-hi gdon/ URL:https://datascience.utah.edu/talks/2022-03-25-chad-topaz-jude-higdon/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1250 | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeFpQN0lB MEZQQ2Z1ZUdFUT09 CATEGORIES:fairness & ethics,statistics END:VEVENT BEGIN:VEVENT UID:2022-04-01-debanjan-mahata@datascience.utah.edu DTSTAMP:20220401T210000Z DTSTART;TZID=America/Denver:20220401T150000 DTEND;TZID=America/Denver:20220401T160000 SUMMARY:Identifying Keyphrases from Text Documents - From Heuristics to Lan guage Models - Debanjan Mahata DESCRIPTION:Debanjan Mahata (Moody Analytics)\n\nAutomatic identification o f keyphrases from text documents is an extreme summarization problem that lies at the intersection of the areas of natural language processing (NLP) and information retrieval (IR). Keyphrases aid in capturing the most sali ent topics from the input text and are useful in multiple downstream tasks such as classification\, clustering\, summarization\, document recommenda tion\, query expansion\, interactive document retrieval\, semantic and fac eted search. Despite the ground-breaking advancements triggered by deep ne ural networks\, automatically identifying keyphrases from text using machi ne learning techniques is still a challenging problem that hasn't been exp lored as much as other related and popular tasks such as named entity extr action\, question answering\, summarization. This talk will provide an ove rview of the advances made in keyphrase extraction and generation from tex t documents and present the approaches\, datasets\, evaluation strategies\ , and the associated challenges. It will also dive into the topic of how l anguage models have been effective in pushing state-of-the-art performance s in this domain. Lastly\, the speaker will present the current trends and future directions for research in this domain.\n\nhttps://datascience.uta h.edu/talks/2022-04-01-debanjan-mahata/ URL:https://datascience.utah.edu/talks/2022-04-01-debanjan-mahata/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:" | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeFpQN0lBMEZQQ2Z 1ZUdFUT09 CATEGORIES:deep learning,machine learning,natural language processing,socie ty & policy END:VEVENT BEGIN:VEVENT UID:2022-04-08-tao-li@datascience.utah.edu DTSTAMP:20220408T210000Z DTSTART;TZID=America/Denver:20220408T150000 DTEND;TZID=America/Denver:20220408T160000 SUMMARY:Improving Data Efficiency of Neural Models using Logic - Tao Li DESCRIPTION:Tao Li (Google Research)\n\nIn this talk\, we will focus on a s imple approach that uses logic to improve neural model performance for nat ural language processing (NLP) tasks. Many downstream NLP tasks involve do main knowledge that can be easily stated in logical forms. We argue that w e can use such knowledge to improve model learning. This results in better data efficiency\, i.e.\, a model that performs better with less annotatio n. To this end\, we propose frameworks that integrate domain knowledge\, e xpressed as declarative constraints\, with neural models. We show that suc h integration substantially improves state-of-the-art neural models in a v ariety of NLP tasks. To facilitate using our frameworks\, we will also pro pose a PyTorch library that unifies differentiable tensor operations and l ogical operations.\n\nhttps://datascience.utah.edu/talks/2022-04-08-tao-li / URL:https://datascience.utah.edu/talks/2022-04-08-tao-li/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:natural language processing END:VEVENT BEGIN:VEVENT UID:2022-04-15-abhinav-kumar@datascience.utah.edu DTSTAMP:20220415T210000Z DTSTART;TZID=America/Denver:20220415T150000 DTEND;TZID=America/Denver:20220415T160000 SUMMARY:Mathematical Modeling for Landmark and 3D Object Detection - Abhina v Kumar DESCRIPTION:Abhinav Kumar\n\nModern computer vision models have excelled on several tasks and beaten several benchmarks. However\, many of these mode ls have avoided the mathematical and principled approaches to computer vis ion\, resulting in sub-optimal performance and issues with interpretabilit y. This talk will re-introduce mathematical modeling for computer vision t asks such as facial landmark localization and monocular 3D object detectio n for autonomous driving. In particular\, we discuss the joint estimation of location\, uncertainty\, and visibility for facial landmark detection a nd introduce a mathematically differentiable NMS for monocular 3D detectio n. The mathematical modeling enables end-to-end learning in these tasks\, resulting in improved performance.\n\nhttps://datascience.utah.edu/talks/2 022-04-15-abhinav-kumar/ URL:https://datascience.utah.edu/talks/2022-04-15-abhinav-kumar/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:computer vision,robotics END:VEVENT BEGIN:VEVENT UID:2022-04-22-khyati-chandu@datascience.utah.edu DTSTAMP:20220422T210000Z DTSTART;TZID=America/Denver:20220422T150000 DTEND;TZID=America/Denver:20220422T160000 SUMMARY:Anchoring Multimodal Narrative Generation - Khyati Chandu DESCRIPTION:Khyati Chandu (Meta AI)\n\nHumans inherently learn from and int eract with multiple views of information\, be it various modalities or lan guages. So\, the expectations from contemporary and 21st-century technolog y are a testimony to the increasing need to model these multiview contexts better. Natural language generation plays a pivotal role in communicating these contexts in human-understandable languages. This talk brings togeth er both of these transformative technologies to make strides toward a long standing dream of human-like multiview narrative generation. The critical challenge is identifying the natural-sounding properties of long-form text s and modeling them in tandem with visual contexts. I present anchors for grounding three such properties including content (relevance)\, structure (coherence)\, and surface form realization (expression)\, and anchors them with relevant visual contexts. These anchors also provide us with human i nterpretable handles for controlling these properties.\n\nDetails: To illu strate the effectiveness of the anchors for each of the three properties\, I present: Starting with content: In situated multimodal contexts\, relev ance is the concept of the elements in one modality being connected to the other modality that makes this context informative and complementary. I p resent visual infilling with curriculum learning as a global objective for content and hierarchically attending over entity skeletons as a local obj ective for content\, to generate visual stories and procedures. To improve the controllability and transferability in English and five other languag es\, I also introduce a dual-stage model with weakly supervised skeletons and a text-as-side attention mechanism to denoise the content in an image caption. Moving onto structure: The alignment of descriptions in language to the corresponding visual inputs is crucial to generating a logical and coherent narrative. I present a scaffolding technique as a local objective for structure by extracting a layout from vast amounts of unsupervised te xt to incorporate structure into cooking recipes generated from images. Fi nally\, surface form: The crux of naturalness to automatic generation come s by incorporating individualized and personalized ways of expressing the same content. I present a locally guided\, weakly supervised model for gen erating persona-based visual stories and a dual-staged adversarial techniq ue to generate mixed views from non-parallel data. All of the above work m ainly focuses on static multimodal narratives\, and finally\, I present a case to highlight the significance of transitioning to dynamic grounding. I conclude by presenting the shortcomings of the current approaches in the NLP domain to the grounding problem and offer recommendations along with executable actions for course correction to bridge this gap and enable gro unding for machines to resemble human communication.\n\nhttps://datascienc e.utah.edu/talks/2022-04-22-khyati-chandu/ URL:https://datascience.utah.edu/talks/2022-04-22-khyati-chandu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1250 and Zoom | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5 GeFpQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:computer vision,education,machine learning,natural language proc essing END:VEVENT BEGIN:VEVENT UID:2022-04-29-james-brundage@datascience.utah.edu DTSTAMP:20220429T210000Z DTSTART;TZID=America/Denver:20220429T150000 DTEND;TZID=America/Denver:20220429T160000 SUMMARY:Leveraging Unlabeled Data for Machine Learning in the Electrocardio gram - James Brundage DESCRIPTION:James Brundage (UU HSC)\n\nSupervised deep learning (DL) has be come an increasingly common tool for advanced analysis of the electrocardi ogram (ECG). These methods rely heavily on labeled datasets\, in which the re is a clinical annotation for each ECG. However\, real world ECG dataset s may not contain enough labeled recordings to facilitate robust feature e xtraction\, preventing DL analysis for clinical problems with small datase ts. Self-supervised learning (SSL) seeks to utilize cheaply labeled or unl abeled data to improve performance in a supervised learning task. This pro cess consists of first training a model on a primary task with cheap data labels\, followed by a second training process which attempts to learn the downstream task by initializing with weights learned from the first. Whil e SSL has become a popular tool in many machine learning domains\, it is o nly starting to be used in ECG based machine learning. Here\, we demonstra te the progress we have made in applying SSL approaches to detect low left ventricular ejection fraction\, a complex ECG detection task\, using data extracted from the University of Utah.\n\nhttps://datascience.utah.edu/ta lks/2022-04-29-james-brundage/ URL:https://datascience.utah.edu/talks/2022-04-29-james-brundage/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) | https://utah.zoom.us/j/94615915833?pwd=TW50Mk5GeF pQN0lBMEZQQ2Z1ZUdFUT09 CATEGORIES:health & medicine,machine learning END:VEVENT BEGIN:VEVENT UID:2022-08-24-bei-wang-phillips@datascience.utah.edu DTSTAMP:20220824T163000Z DTSTART;TZID=America/Denver:20220824T103000 DTEND;TZID=America/Denver:20220824T114500 SUMMARY:On Hypergraph Analysis and Visualization - Bei Wang Phillips DESCRIPTION:Bei Wang Phillips (Utah SoC/SCI)\n\nHypergraphs capture multi-w ay relationships in data\, and they have\nconsequently seen a number of ap plications in higher-order network\nanalysis\, computer vision\, geometry processing\, and machine learning.\nIn this talk\, I will discuss hypergra ph analysis and visualization.\nIn particular\, I will focus on developing the theoretical foundations\nin studying the space\nof hypergraphs using ingredients from optimal transport.\n\nThis talk is\nbased on joint works\ nwith Youjia Zhou\, Archit Rathore\, Emilie Purvine\, Samir Chowdhury\, To m\nNeedham\, and Ethan Semrad.\n\nhttps://datascience.utah.edu/talks/2022- 08-24-bei-wang-phillips/ URL:https://datascience.utah.edu/talks/2022-08-24-bei-wang-phillips/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:networks & graphs,visualization END:VEVENT BEGIN:VEVENT UID:2022-08-31-casey-greene@datascience.utah.edu DTSTAMP:20220831T163000Z DTSTART;TZID=America/Denver:20220831T103000 DTEND;TZID=America/Denver:20220831T114500 SUMMARY:Talk by Casey Greene - Casey Greene DESCRIPTION:Casey Greene (CU Anschutz)\n\nhttps://datascience.utah.edu/talk s/2022-08-31-casey-greene/ URL:https://datascience.utah.edu/talks/2022-08-31-casey-greene/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 END:VEVENT BEGIN:VEVENT UID:2022-09-07-kevin-moon@datascience.utah.edu DTSTAMP:20220907T163000Z DTSTART;TZID=America/Denver:20220907T103000 DTEND;TZID=America/Denver:20220907T114500 SUMMARY:Scalable supervised manifold learning with random forests and neura l networks - Kevin Moon DESCRIPTION:Kevin Moon (USU)\n\nThe manifold assumption has been used in ma ny machine learning applications to combat the curse of dimensionality. Mo st manifold learning methods are unsupervised and typically focus on prese rving the dominant structure and variation in the data. In many cases\, we wish to analyze the data in a supervised setting with respect to expert-p rovided data labels. Most supervised manifold learning methods exaggerate the separation between data points of different classes\, distorting the t rue structure of the data. In this talk\, I will present RF-PHATE\, a supe rvised dimensionality reduction method that preserves the true structure o f the variables that are relevant for the supervised task. RF-PHATE is bas ed upon a diffusion process applied to random forest proximities and is we ll-suited for data visualization. I will then show how to improve the scal ability of RF-PHATE and any other manifold learning algorithm and perform out of sample extension using geometry regularized autoencoders (GRAE).\n\ nhttps://datascience.utah.edu/talks/2022-09-07-kevin-moon/ URL:https://datascience.utah.edu/talks/2022-09-07-kevin-moon/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:deep learning,machine learning,visualization END:VEVENT BEGIN:VEVENT UID:2022-09-14-elliot-smith@datascience.utah.edu DTSTAMP:20220914T163000Z DTSTART;TZID=America/Denver:20220914T103000 DTEND;TZID=America/Denver:20220914T114500 SUMMARY:Human neuronal population encoding of temporal difference learning variables during risky choices - Elliot Smith DESCRIPTION:Elliot Smith (Utah Neurology)\n\nRecent research in AI showed t hat agents designed to predict the full distribution of potential rewards\ , rather than a central estimate of that distribution\, generate richer le arning distributions that allow them to perform better\, especially on ris ky tasks. Such distributional reinforcement learning (distRL) was also dis covered in dopamine neurons in the rodent ventral tegmental area. In this nanosymposium presentation\, I will discuss recent work from direct brain recordings in neurosurgical patients undergoing monitoring for treatment o f medically refractory epilepsy who performed a risky decision making task called the Balloon Analog Risk Task. Results from two studies will be pre sented: In the first study\, we examined neuronal population recordings (1 57 neurons) from microelectrodes implanted in the anterior cingulate\, orb itofrontal and temporal cortices (15 participants)\, finding that human pr efrontal and mesial temporal neurons exhibited signatures of distRL: corre lated diverse optimism in reward coding and diverse asymmetric scaling of reward prediction error. In the second study\, we examined correlations be tween broadband high frequency local field potentials (an established corr elate of population neuronal firing) and variables from temporal differenc e learning models for reward and risk while 37 participants made risky cho ices during BART. We found differences in which brain areas (3199 stereoel ectroencephalography or electrocorticography contacts sampling frontal\, t emporal\, and parietal lobes) encoded temporal difference learning model v ariables between participants who were more or less risk averse in their c hoices during BART. These areas included the left dorsolateral prefrontal\ , anterior cingulate\, and orbitofrontal cortices. The results from these studies shed light on the neural underpinnings of human value learning in uncertain environments.\n\nhttps://datascience.utah.edu/talks/2022-09-14-e lliot-smith/ URL:https://datascience.utah.edu/talks/2022-09-14-elliot-smith/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:biology & genomics,health & medicine,human-centered computing END:VEVENT BEGIN:VEVENT UID:2022-09-21-jes-ford@datascience.utah.edu DTSTAMP:20220921T163000Z DTSTART;TZID=America/Denver:20220921T103000 DTEND;TZID=America/Denver:20220921T114500 SUMMARY:Model Review: Improving Transparency\, Reproducibility\, & Knowledg e Sharing using MLflow - Jes Ford DESCRIPTION:Jes Ford (Cash App)\n\nCode Review is an integral part of softw are development\, but many teams don’t have similar processes in place f or the development and deployment of Machine Learning (ML) models. I will motivate the decision to create a Model Review process\, starting from the principles of transparency\, reproducibility\, and knowledge sharing. MLf low is a useful Python package to help simplify and automate much of the t racking necessary to create detailed records of machine learning experimen ts. Much of this talk will be spent introducing this tool\, and demonstrat ing the core MLflow Tracking functionality. I’ll discuss how my team is currently running a Model Review process for any ML models that we push to production\, and how we use MLflow to streamline this work and learn from each other.\n\nhttps://datascience.utah.edu/talks/2022-09-21-jes-ford/ URL:https://datascience.utah.edu/talks/2022-09-21-jes-ford/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:machine learning END:VEVENT BEGIN:VEVENT UID:2022-09-28-prashant-pandey@datascience.utah.edu DTSTAMP:20220928T163000Z DTSTART;TZID=America/Denver:20220928T103000 DTEND;TZID=America/Denver:20220928T114500 SUMMARY:Scalability Challenges in Large-Scale Sequence Search - Prashant Pa ndey DESCRIPTION:Prashant Pandey (Utah SoC)\n\nSequence-level searches on large collections of RNA sequencing experiments\, such as the NCBI Sequence Read Archive (SRA)\, would enable one to ask many questions about the expressi on or variation of a given transcript in a population. Building an efficie nt and scalable sequence search index at the scale of SRA data is a challe nging task and requires fundamental innovations in compression and scalabl e indexing. Recently\, several tools have been proposed to index and searc h through SRA data but they offer various trade-offs in terms of space\, s peed\, updatability\, and accuracy. In this talk\, I will present Mantis\, a fast\, exact\, and updatable sequence search index. Mantis uses recent advancements in fast and compact hash tables\, domain-specific data compre ssion techniques\, and scalable indexing to build a scalable and updatable index and supports fast sequence searches on ~40K experiments (>100TB) in size from SRA.\n\nhttps://datascience.utah.edu/talks/2022-09-28-prashant- pandey/ URL:https://datascience.utah.edu/talks/2022-09-28-prashant-pandey/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:biology & genomics,data management END:VEVENT BEGIN:VEVENT UID:2022-10-05-jessica-shi@datascience.utah.edu DTSTAMP:20221005T163000Z DTSTART;TZID=America/Denver:20221005T103000 DTEND;TZID=America/Denver:20221005T114500 SUMMARY:Theoretically and Practically Efficient Parallel Nucleus Decomposit ion - Jessica Shi DESCRIPTION:Jessica Shi (MIT)\n\nWe study the nucleus decomposition problem \, which has been shown to be useful in finding dense substructures in gra phs. We present a novel parallel algorithm that is efficient both in theor y and in practice. Our algorithm achieves a work complexity matching the b est sequential algorithm while also having low depth (parallel running tim e)\, which significantly improves upon the only existing parallel nucleus decomposition algorithm (Sariyuce et al.\, PVLDB 2018). The key to the the oretical efficiency of our algorithm is a new lemma that bounds the amount of work done when peeling cliques from the graph\, combined with the use of theoretically-efficient parallel algorithms for clique listing and buck eting.\n\nWe introduce several new practical optimizations\, including a n ew multi-level hash table structure to store information on cliques space- efficiently and a technique for traversing this structure cache-efficientl y. On a 30-core machine with two-way hyper-threading on real-world graphs\ , we achieve up to a 55x speedup over the state-of-the-art parallel nucleu s decomposition algorithm by Sariyuce et al.\, and up to a 40x self-relati ve parallel speedup. We are able to efficiently compute larger nucleus dec ompositions than prior work on several million-scale graphs for the first time.\n\nhttps://datascience.utah.edu/talks/2022-10-05-jessica-shi/ URL:https://datascience.utah.edu/talks/2022-10-05-jessica-shi/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:algorithms & theory,education,networks & graphs END:VEVENT BEGIN:VEVENT UID:2022-10-19-jie-zhang@datascience.utah.edu DTSTAMP:20221019T163000Z DTSTART;TZID=America/Denver:20221019T103000 DTEND;TZID=America/Denver:20221019T114500 SUMMARY:Active Sampling for Min-Max Fairness - Jie Zhang DESCRIPTION:Jie Zhang (U Washington)\n\nModels satisfying Min-max fairness minimizes maximum group specific losses\, so that the model has a more equ itable performance over all groups. Benefit of min-max fair models over ot her fairness notions and models include it levels up: meaning it only degr ades performance of a group if the degradation improves performance on the worst off group.\n\nIn this talk\, I will briefly mention the prior works that defined min-max fairness. I will mainly focus on our paper “Active Sampling for Min-Max Fairness” to introduce two algorithms that find mi n-max fair models with convergence guarantees.\n\nhttps://datascience.utah .edu/talks/2022-10-19-jie-zhang/ URL:https://datascience.utah.edu/talks/2022-10-19-jie-zhang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4R0RiSUpsb3k vdz09 CATEGORIES:fairness & ethics,statistics END:VEVENT BEGIN:VEVENT UID:2022-11-02-alex-chin@datascience.utah.edu DTSTAMP:20221102T163000Z DTSTART;TZID=America/Denver:20221102T103000 DTEND;TZID=America/Denver:20221102T114500 SUMMARY:Quantifying supply-demand imbalance in ridesharing systems - Alex C hin DESCRIPTION:Alex Chin (Lyft)\n\nThe status of the rider and driver distribu tions and how they interact in a two-sided marketplace has implications fo r market efficiency and policy-making. As such\, accurately characterizing the supply-demand state of Lyft's marketplace is a crucial task. I will d escribe how this task can be framed in terms of an asymmetric optimal tran sport problem that yields a multi-resolution view of the supply-demand sta te and how it varies both spatially and temporally. I will then discuss ho w this approach can be incorporated into applications such as policy optim ization and machine learning prediction problems. Time permitting I will a lso discuss other science efforts we have at Lyft.\n\nhttps://datascience. utah.edu/talks/2022-11-02-alex-chin/ URL:https://datascience.utah.edu/talks/2022-11-02-alex-chin/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:society & policy END:VEVENT BEGIN:VEVENT UID:2022-11-09-shireen-elhabian@datascience.utah.edu DTSTAMP:20221109T173000Z DTSTART;TZID=America/Denver:20221109T103000 DTEND;TZID=America/Denver:20221109T114500 SUMMARY:Data-driven Shape Analysis: Methods\, Applications\, and Future - S hireen Elhabian DESCRIPTION:Shireen Elhabian (Utah CS/SCI)\n\nQuantitative analysis of shap es is contingent upon defining a metric in the space of shapes to compare shapes and perform shape statistics. A growing consensus in the field that such a metric should be adapted to the specific population under investig ation\, begging for learning such a metric in a data-driven manner. This t alk will cover a state-of-art data-driven approach for statistical shape m odeling that provides unbiased\, objective\, and intuitive evaluation of g eometric shapes\, emphasizing anatomical structures reconstructed from vol umetric images. I will talk about how we applied shape analysis to clinica l and scientific questions in various ways. I will also highlight the role of machine learning in mitigating critical bottlenecks and significant ba rriers to making shape modeling a robust tool for on-demand clinical diagn ostics and streamlining its adoption in research and practice. I will end with a future outlook for shape modeling to enable more complex and divers e modeling scenarios.\n\nhttps://datascience.utah.edu/talks/2022-11-09-shi reen-elhabian/ URL:https://datascience.utah.edu/talks/2022-11-09-shireen-elhabian/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:health & medicine,statistics END:VEVENT BEGIN:VEVENT UID:2022-11-16-bernadette-stolz@datascience.utah.edu DTSTAMP:20221116T173000Z DTSTART;TZID=America/Denver:20221116T103000 DTEND;TZID=America/Denver:20221116T114500 SUMMARY:Applications of global and local persistent homology for the shape of biological data - Bernadette Stolz DESCRIPTION:Bernadette Stolz (Oxford)\n\nIn the first part of this talk\, I will showcase how persistent homology can be used to spatially characteri se structural abnormality in tumour blood vessel networks. More specifical ly\, I will show that the number of vessel loops and their distribution in these networks change over time when tumours undergo treatment with vascu lar targeting agents and radiation therapy. In the second part of the talk \, I will speak about applications of local persistent homology. I will sh ow how local persistent homology can be used to select landmarks from larg e and noisy data sets. In contrast to existing methods\, this subsampling process is robust to outliers and is developed specifically for persistent homology. I will further introduce a novel method that can detect geometr ic anomalies\, such as intersections or boundaries\, in point cloud data s ampled from intersecting surfaces. This detection is based on the computat ion of persistent homology in local annular neighbourhoods around points a nd is less sensitive to the size of the local neighbourhood and surface cu rvature than an existing method.\n\nhttps://datascience.utah.edu/talks/202 2-11-16-bernadette-stolz/ URL:https://datascience.utah.edu/talks/2022-11-16-bernadette-stolz/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4R0RiSUpsb3k vdz09 CATEGORIES:biology & genomics END:VEVENT BEGIN:VEVENT UID:2022-11-30-aaron-quinlan@datascience.utah.edu DTSTAMP:20221130T173000Z DTSTART;TZID=America/Denver:20221130T103000 DTEND;TZID=America/Denver:20221130T114500 SUMMARY:Finding signals of genome mutation in the noise of DNA sequencing e rror - Aaron Quinlan DESCRIPTION:Aaron Quinlan (Utah Human Genetics)\n\nThe research in our labo ratory is focused on the application of computational methods to develop a deeper understanding of genetic variation in diverse contexts. Modern exp erimental methods allow us to examine entire genomes with exquisite detail . Perhaps not surprisingly\, staggering complexity is revealed as we look more closely at how genetic variation (both inherited and somatic) contrib utes to phenotypes. Modern genomic technologies necessitate efficient appr oaches for exploring\, manipulating and comparing large genomic datasets. We develop such methods so that we and others may apply them to experiment s investigating the impact of genetic variation on human disease\, evoluti on\, and somatic differentiation. Genome research is difficult - we strive to develop computational means that make it easier.\n\nhttps://datascienc e.utah.edu/talks/2022-11-30-aaron-quinlan/ URL:https://datascience.utah.edu/talks/2022-11-30-aaron-quinlan/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:biology & genomics END:VEVENT BEGIN:VEVENT UID:2022-12-07-tao-yang@datascience.utah.edu DTSTAMP:20221207T173000Z DTSTART;TZID=America/Denver:20221207T103000 DTEND;TZID=America/Denver:20221207T114500 SUMMARY:Optimizing Ranking Effectiveness and Fairness - Tao Yang DESCRIPTION:Tao Yang (Utah SoC)\n\nAdvanced ranking techniques have led to improvements in AI-powered information services that significantly changed people's lives. For example\, search engines that rank information accord ing to their utilities to use's queries have helped billions of people bet ter finish their tasks in daily work\; recommendation systems that rank pr oducts/movies/news according to the user's interests have completely chang ed the way people discover information everyday. Therefore\, how to constr uct and optimize ranking systems is one of the most important research pro blems in the field of Information Retrieval (IR). When optimizing ranking systems\, there are two important criteria to measure the quality of resul t rankings in IR systems. The first criterion is ranking effectiveness\, w hich refers to the ability of a ranking system to effectively present resu lts based on their relevance to the users' needs. The second criterion is ranking fairness\, which refers to the ability of a ranking system to pres ent results fairly. For example\, in job recommendation\, if a ranking sys tem only considers ranking effectiveness and ranks items solely according to relevance\, a small number of top candidates will always be exposed to users and dominate users' attention as users usually only examine the top ranks. In such case\, other candidates will rarely have the chance to be h ired even when they are highly qualified for the job. Therefore\, it is im portant to balance the effectiveness of ranked lists with the fairness in ranking optimization.\nIn this talk\, I will present my recent works on ra nking effectiveness and fairness optimization. The talk will be two parts. In the first part of this talk\, I will introduce works on sole-effective ness optimization where I propose uncertainty-aware rank systems based on Bayes modelling. In the second part of this talk\, I will introduce works on fairness-effectiveness joint optimization.\n\nhttps://datascience.utah. edu/talks/2022-12-07-tao-yang/ URL:https://datascience.utah.edu/talks/2022-12-07-tao-yang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:fairness & ethics,optimization END:VEVENT BEGIN:VEVENT UID:2023-01-11-george-vega-yon@datascience.utah.edu DTSTAMP:20230111T174500Z DTSTART;TZID=America/Denver:20230111T104500 DTEND;TZID=America/Denver:20230111T114500 SUMMARY:Prediction of Gene Functions by Leveraging Biological Insights with Mechanistic Machine Learning - George Vega Yon DESCRIPTION:George Vega Yon (Utah Epidemiology)\n\nBiomedical sciences\, in particular\, bioinformaticians and computational biologists\, are in a ra ce to annotate the immense number of genes and gene products we are still learning from. In this talk\, I present a new method for predicting gene f unctions using mechanistic machine learning. Mechanistic machine learning is a relatively new area of research where predictive ML-based algorithms are improved by incorporating domain knowledge via mechanistic models. Her e\, we use a theoretically-funded function evolution model that relies sol ely on phylogenetic trees from PantherDB and annotations from the Gene Ont ology (GO) to make high-quality predictions. I will illustrate how combini ng the mentioned model with a large gene expression database (Bgee) in an ML model significantly improves prediction quality.\n\nhttps://datascience .utah.edu/talks/2023-01-11-george-vega-yon/ URL:https://datascience.utah.edu/talks/2023-01-11-george-vega-yon/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:biology & genomics,machine learning,networks & graphs,statistics END:VEVENT BEGIN:VEVENT UID:2023-01-18-echo-warner@datascience.utah.edu DTSTAMP:20230118T174500Z DTSTART;TZID=America/Denver:20230118T104500 DTEND;TZID=America/Denver:20230118T120000 SUMMARY:College of Nursing\, Division of Acute and Chronic Care - Echo Warn er DESCRIPTION:Echo Warner (Utah Nursing\, HCI)\n\nUnproven health claims on t he internet may have substantial influence on patient behaviors and decisi on making. For example\, cancer patients with curable disease who pursue u nproven cancer treatment in lieu of evidence-based approaches demonstrate 2-4 times higher mortality than patients who avoid unproven cancer treatme nt. Interventions to mitigate the impact of online health misinformation a re desperately needed. The few interventions that have been tested do not consider the extent of exposure to misinformation online\, primarily becau se individual estimates of exposure are based on self-report and are consi dered highly unreliable. Web-monitoring software is typically used by busi nesses to monitor remote employee productivity. Our paradigm shifting appr oach applies web-monitoring to quantify online cancer misinformation expos ure. We will discuss the feasibility of using web-monitoring software to q uantify exposure to online health information with a special focus on canc er symptom management and unproven cancer treatment misinformation. Specif ically\, we will review 1) characteristics of online cancer health misinfo rmation and the impacts this exposure may have on cancer patient health ou tcomes\, relationships\, and finances 2) Ethical considerations of web-mon itoring\, and 3) methodological rigor and reproducibility of web-monitorin g approaches for studying exposure to other types of health misinformation online.\n\nhttps://datascience.utah.edu/talks/2023-01-18-echo-warner/ URL:https://datascience.utah.edu/talks/2023-01-18-echo-warner/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Virtual: | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:health & medicine,society & policy END:VEVENT BEGIN:VEVENT UID:2023-01-25-shweta-jain@datascience.utah.edu DTSTAMP:20230125T174500Z DTSTART;TZID=America/Denver:20230125T104500 DTEND;TZID=America/Denver:20230125T114500 SUMMARY:Putting Parameterization into Practice - Shweta Jain DESCRIPTION:Shweta Jain (Utah SoC)\n\nGraphs are everywhere: social network s\, protein interaction networks\, citation networks\, epidemic spread net works. Applications that use these graphs have progressively become more s ophisticated\, relying on getting fast and accurate solutions to graph-the oretic problems\, many of which are NP-Hard. However\, the sizes of today' s graphs easily run into millions of vertices and edges\, if not more. As a result\, many classic algorithms are infeasible for such graphs.\n\nFort unately\, real-world graphs across different domains show a lot of common characteristics such as an abundance of triangles\, low average distance b etween vertices (small-world property)\, low degeneracy (a measure of the sparsity of edges) etc. which we can leverage to design algorithms that ar e provably efficient given those parameters. I will demonstrate this in th e context of clique counting and decomposition of graphs\, which has appli cations in community detection\, spam detection\, fraud detection\, gene m odule detection etc. The new algorithms not only massively improve the per formance in practice but also improve our theoretical understanding of rea l-world graphs and help to bridge the gap between theory and practice.\n\n https://datascience.utah.edu/talks/2023-01-25-shweta-jain/ URL:https://datascience.utah.edu/talks/2023-01-25-shweta-jain/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:algorithms & theory,biology & genomics,networks & graphs END:VEVENT BEGIN:VEVENT UID:2023-02-01-ana-marsovic@datascience.utah.edu DTSTAMP:20230201T174500Z DTSTART;TZID=America/Denver:20230201T104500 DTEND;TZID=America/Denver:20230201T114500 SUMMARY:“AI” That Masters Language Could Reason about Negation - Ana Ma rsovic DESCRIPTION:Ana Marsovic (Utah SoC)\n\nWe have experienced firsthand the gr owing impact that AI technologies like ChatGPT have. It is agreed upon tha t risks involving such technologies must be managed. This is at a glance a kin to how people handle safety-critical systems in\, e.g.\, aviation. How ever\, the rigorous principles of safety engineering are not easily applic able to AI technologies. AI-backed solutions are obscured in AI’s intern als\, and the exact requirements for AI safety are unverifiable since they are neither defined nor regulated. One technique that has emerged as a st ep to measure an aspect of AI safety is to construct data that represents a specific phenomenon that a safe model must handle well\, and that has no spurious correlations. Low performance on the dataset is undesired.In thi s talk\, I’ll focus on one common linguistic phenomenon\, negation\, wit hout which the full power of human language-based communication cannot be realized. I will show how we carefully constructed a question-answering da taset\, CondaQA\, to study how well current models (described by the New Y ork Time Magazine as “mastering language”) reason about negated statem ents. An InstructGPT model (the latest GPT model before Nov 28\, 2022) com bined with chain-of-thought prompting achieves 66.28% accuracy and 27.28% consistency on CondaQA\, way behind human accuracy of 91.94% and consisten cy of 81.58%.\n\nhttps://datascience.utah.edu/talks/2023-02-01-ana-marsovi c/ URL:https://datascience.utah.edu/talks/2023-02-01-ana-marsovic/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:large language models,natural language processing END:VEVENT BEGIN:VEVENT UID:2023-02-08-aaron-clauset@datascience.utah.edu DTSTAMP:20230208T174500Z DTSTART;TZID=America/Denver:20230208T104500 DTEND;TZID=America/Denver:20230208T114500 SUMMARY:Meritocracy or systemic bias? Untangling the drivers the productivi ty and prominence among scientists - Aaron Clauset DESCRIPTION:Aaron Clauset (UC Boulder)\n\nSimple measures of scholarly prod uctivity and prominence vary enormously across both individual scientists and institutions -- but to what degree do these inequalities represent gen uine meritocratic differences vs. systemic biases that limit scientific pr ogress?\n\nIn this talk\, I'll describe a sequence of results that substan tially untangle the underlying systemic drivers of productivity and promin ence among scientists. First\, I'll show that productivity and prominence are\, to a significant degree\, environmental variables such that the pres tige of a scientist's working environment drives their individual producti vity\, largely by providing larger research groups to elite scientists. Se cond\, I'll describe a network-based generative model of individual produc tivity and prominence that untangles these measures from their underlying collaboration networks. These models corroborate the labor-advantage hypot hesis of elite institutions\, and also reveal both that gendered differenc es in the productivity and prominence of mid-career researchers can be lar gely explained by gendered differences in coauthorship networks\, and that these networks are partially transferable from senior to junior collabora tors. Hence\, collaboration networks\, and the systemic factors that shape them\, play a critical role in driving scholarly inequalities in science\ , and suggest that these networks are an important form of unequally distr ibuted social capital that shapes who makes what scientific discoveries. I 'll close with a discussion of policies that could potentially mitigate th e unequal distribution of this social capital and help both diversify the academy and broaden its contributions to society.\n\nhttps://datascience.u tah.edu/talks/2023-02-08-aaron-clauset/ URL:https://datascience.utah.edu/talks/2023-02-08-aaron-clauset/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:fairness & ethics,networks & graphs,society & policy END:VEVENT BEGIN:VEVENT UID:2023-02-15-orly-alter@datascience.utah.edu DTSTAMP:20230215T174500Z DTSTART;TZID=America/Denver:20230215T104500 DTEND;TZID=America/Denver:20230215T114500 SUMMARY:Solving Cancer with Data: Mathematical Discovery and Computational and Experimental Validation of Whole-Genome Genotype–Survival and Respon se to Treatment Phenotype Relationships in Cancer - Orly Alter DESCRIPTION:Orly Alter (Utah BME\, SCI)\n\n1/2 of men and 1/3 of women will face cancer\, a disease of the whole 3B-nucleotide genome. But\, despite the availability of open-source data and the $100/1-hour genome\, genetic tests remain limited to one to a few hundred genes. Therefore\, the progno sis\, diagnosis\, and treatment of cancer remain unchanged. This is due to the lack of suitable AI/ML. I will describe work in my lab inventing AI/M L that connects the whole genome with a patient’s survival and response to treatment. Our algorithms discover accurate\, precise\, and interpretab le predictors\, applicable to the general population\, from as few as 50 –100 patients. Our predictors outperform all other indicators\, where th ey exist. All other methods miss them. I will describe my international re trospective clinical trial\, which validated a genome-wide pattern in tumo rs from glioblastoma brain cancer patients as the best predictor of life e xpectancy and response to standard of care. We discovered this\, and predi ctors in\, e.g.\, adult lung\, ovarian\, and uterine adenocarcinoma tumors and pediatric nerve neuroblastoma tumors\, in public data\, proving that the algorithms and predictors are uniquely suited to personalized medicine . I will also describe work translating the algorithms and predictors to t he clinic.\n\nhttps://datascience.utah.edu/talks/2023-02-15-orly-alter/ URL:https://datascience.utah.edu/talks/2023-02-15-orly-alter/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:algorithms & theory,biology & genomics,health & medicine END:VEVENT BEGIN:VEVENT UID:2023-02-22-vivek-gupta@datascience.utah.edu DTSTAMP:20230222T174500Z DTSTART;TZID=America/Denver:20230222T104500 DTEND;TZID=America/Denver:20230222T114500 SUMMARY:Inference and Reasoning for Semi-structured Tables - Vivek Gupta DESCRIPTION:Vivek Gupta (Utah SoC)\n\nUnderstanding semi-structured tabular data\, which is ubiquitous in the real world\, requires an understanding of the meaning of text fragments and the implicit connections between them . We believe such data could be used to investigate how individuals and ma chines reason about semi-structured data. First\, we present the InfoTabS dataset\, which consists of human-written textual predictions based on tab les collected from Wikipedia's infoboxes. Our research demonstrates that t he semi-structured\, multi-domain\, and heterogeneous nature of the premis es prompts complicated\, multi-faceted reasoning\, offering a modeling cha llenge for traditional modeling techniques. Second\, we analyzed these cha llenges in-depth and developed simple\, effective preprocessing strategies to overcome them. Thirdly\, despite accurate NLI prediction\, we demonstr ate through rigorous probing that the existing model does not reason with the provided tabular facts. To address this\, we suggest a two-stage evide nce extraction and tabular inference technique for enhancing model reasoni ng and interpretability. We also investigate efficient methods for enhanci ng tabular inference datasets with semi-automatic data augmentation and pa ttern-based pre-training. Lastly\, to ensure that tabular reasoning models work in more than one language\, we introduce XInfoTabS\, a unique proble m of bilingual tabular inference\, and a cost-effective pipeline for trans lating tables. In the near future\, we plan to test the tabular reasoning model for temporal changes\, especially for dynamic tables where informati on changes over time.\n\nhttps://datascience.utah.edu/talks/2023-02-22-viv ek-gupta/ URL:https://datascience.utah.edu/talks/2023-02-22-vivek-gupta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:data management,machine learning,natural language processing END:VEVENT BEGIN:VEVENT UID:2023-03-01-emily-hadley@datascience.utah.edu DTSTAMP:20230301T174500Z DTSTART;TZID=America/Denver:20230301T104500 DTEND;TZID=America/Denver:20230301T114500 SUMMARY:Applied Strategies for Advancing Racial Equity and Addressing Bias in Big Data Research - Emily Hadley DESCRIPTION:Emily Hadley (RTI International)\n\nIn the last decade\, big da ta research studies have proliferated and\, in some cases\, offered consid erable promise for humanity. Yet\, numerous incidents have documented that without safeguards\, this same research can reproduce and amplify existin g societal biases and disparities. Given the increased public awareness of structural racism\, bias based on race and ethnicity in big data research is particularly concerning. Big data researchers can mitigate the risk of perpetuating bias based on race and ethnicity by intentionally incorporat ing best practices that reduce or eliminate racial biases. We synthesize k ey findings and applied recommendations from over 140 sources for addressi ng race and ethnicity bias in big data research. We discuss considerations when planning a big data project\, including identifying study motivation s\, centering participatory involvement\, and addressing key concerns rega rding race and ethnicity in big data collection and quality. We detail iss ues related to proxy discrimination\, algorithmic audits\, data completene ss\, and the Big Data Paradox. We provide real-world examples of advancing racial equity and addressing bias and share recommendations for additiona l resources and opportunities for further investigation.\n\nhttps://datasc ience.utah.edu/talks/2023-03-01-emily-hadley/ URL:https://datascience.utah.edu/talks/2023-03-01-emily-hadley/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:fairness & ethics,society & policy END:VEVENT BEGIN:VEVENT UID:2023-03-15-nate-veldt@datascience.utah.edu DTSTAMP:20230315T164500Z DTSTART;TZID=America/Denver:20230315T104500 DTEND;TZID=America/Denver:20230315T114500 SUMMARY:Measuring homophily in group interactions: hypergraph models and co mbinatorial impossibilities - Nate Veldt DESCRIPTION:Nate Veldt (Texas A&M)\n\nHomophily is the well-known sociologi cal principle that people tend to connect with others who are similar to t hem\, or more informally: "birds of a feather flock together." Although ma ny social interactions occur in groups\, homophily is typically measured u sing a graph\, which only accounts for interactions involving two individu als. This talk will present a new hypergraph framework for more directly m easuring homophily in group settings. Our measures highlight natural patte rns in group homophily that appear with gender in scientific collaboration and political affiliation in legislative bill cosponsorship\, and also re veal distinctive gender distributions in group photographs\, all of which cannot be fully captured by graph-based measures. We will also discuss sub tle combinatorial limits and impossibilities that arise when measuring hom ophily in hypergraphs\, which are completely independent of human behavior and must be properly accounted for in order to understand how homophily c an (and cannot) be manifested in group interactions.\n\nhttps://datascienc e.utah.edu/talks/2023-03-15-nate-veldt/ URL:https://datascience.utah.edu/talks/2023-03-15-nate-veldt/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:networks & graphs END:VEVENT BEGIN:VEVENT UID:2023-03-22-bailey-fosdick@datascience.utah.edu DTSTAMP:20230322T164500Z DTSTART;TZID=America/Denver:20230322T104500 DTEND;TZID=America/Denver:20230322T114500 SUMMARY:Modeling Infection Fatality Rates to Assess the Burden of COVID-19 in Developing Countries - Bailey Fosdick DESCRIPTION:Bailey Fosdick (CU Anschutz)\n\nCOVID-19 spread quickly around the world after first being discovered in China in late 2019. It has had d evastating impacts\, however its impacts\, both in terms of infection prev alence and fatalities\, have been non-uniformly distributed worldwide. Whi le early studies focused on COVID-19 infection and fatality rates in high- income countries\, much less attention has been given to the impacts of CO VID-19 in developing countries. In this work\, we systematically reviewed the literature to identify all COVID-19 serology studies conducted by earl y 2021 using population representative samples. We developed a Bayesian hi erarchical model for simultaneously modeling serology and death data to ma ke inference on age-specific seroprevalence and age-specific infection fat ality rates. This model directly accounts for conventional sampling uncert ainty\, as well as uncertainty about the serological test assay sensitivit y and specificity. Through a careful analysis of data from over thirty dev eloping countries\, we found seroprevalence in many developing country loc ations was markedly higher than in high-income countries early in the pand emic and age-specific infection fatality rates were roughly twice as high as that in high-income countries.\n\nhttps://datascience.utah.edu/talks/20 23-03-22-bailey-fosdick/ URL:https://datascience.utah.edu/talks/2023-03-22-bailey-fosdick/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:statistics END:VEVENT BEGIN:VEVENT UID:2023-03-29-justin-baker@datascience.utah.edu DTSTAMP:20230329T164500Z DTSTART;TZID=America/Denver:20230329T104500 DTEND;TZID=America/Denver:20230329T114500 SUMMARY:Monotone Implicit Graph Neural Networks for Long-Range Dependency L earning - Justin Baker DESCRIPTION:Justin Baker (Utah Math & SCI)\n\nFrom social networks to chemi cal engineering\, deep graph neural networks play an instrumental role in advancing our industrial and scientific frontier. Of particular interest a re networks which can learn long range dependencies in a scalable and expr essive manner. In this talk\, we will delve into the power of deep learnin g on graphs\, with a focus on implicit graph neural networks (IGNNs) and t heir scalability. We will also discuss how monotone operator theory enhanc es the expressivity of IGNNs\, overcoming a crucial obstacle to learning l ong-range dependencies. By doing so\, monotone IGNNs can significantly imp rove graph learning and have the potential to make breakthroughs in variou s fields.\n\nhttps://datascience.utah.edu/talks/2023-03-29-justin-baker/ URL:https://datascience.utah.edu/talks/2023-03-29-justin-baker/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4R0RiSUpsb3k vdz09 CATEGORIES:deep learning,networks & graphs END:VEVENT BEGIN:VEVENT UID:2023-04-05-yao-yaun-mao@datascience.utah.edu DTSTAMP:20230405T164500Z DTSTART;TZID=America/Denver:20230405T104500 DTEND;TZID=America/Denver:20230405T114500 SUMMARY:Searching for dwarf (small) galaxies in astronomical surveys - Yao- Yaun Mao DESCRIPTION:Yao-Yaun Mao (Utah Astro)\n\nDwarf galaxies are small fuzzy gal axies that contain much fewer stars than Milky Way-mass galaxies. Observat ions of these little galaxies can enhance our understanding of galaxy form ation and the nature of dark matter. Finding these dwarf galaxies is\, how ever\, not an easy task because they are faint and dim from our perspectiv e. I will describe the challenges and recent efforts on the search for nea rby dwarf galaxies in astronomical surveys. One particular challenge is to identify potential dwarf galaxies with only image data that do not contai n distance information. I will discuss a few traditional methods that are used to identify dwarf galaxies and obtain their astronomical distances\, and why these methods are mostly used to find dwarf galaxies in specific p atches of the sky. I will then discuss how the distance information can be used as training data in machine learning algorithms\, such as convolutio nal neural networks\, to derive distance information from just astronomica l images. Finally\, I will discuss the remaining challenges we have\, and how we may improve the methods in preparation for future observations from the Rubin Observatory Legacy Survey of Space and Time (LSST) and Roman Sp ace Telescope.\n\nhttps://datascience.utah.edu/talks/2023-04-05-yao-yaun-m ao/ URL:https://datascience.utah.edu/talks/2023-04-05-yao-yaun-mao/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:deep learning,machine learning,physics & astronomy END:VEVENT BEGIN:VEVENT UID:2023-04-12-titus-brown@datascience.utah.edu DTSTAMP:20230412T164500Z DTSTART;TZID=America/Denver:20230412T104500 DTEND;TZID=America/Denver:20230412T114500 SUMMARY:Is everything everywhere all at once? Asking questions of all publi c microbiome shotgun data - Titus Brown DESCRIPTION:Titus Brown (UC Davis)\n\nPublic sequence data offers many oppo rtunities for reuse\, exploration\, and discovery. What happens if you mak e it really\, really easy to search the content of all the public microbio me data sets? It turns out you can enable some interesting science\, but y ou also need to address many technical\, social\, and policy issues. In th is talk I’ll showcase some of our results from being able to search ever ything\, everywhere\, all at once\; describe our current efforts\; and dis cuss the possible opportunities and challenges of petabyte-scale sequence search.\n\nhttps://datascience.utah.edu/talks/2023-04-12-titus-brown/ URL:https://datascience.utah.edu/talks/2023-04-12-titus-brown/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 | https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4 R0RiSUpsb3kvdz09 CATEGORIES:health & medicine,society & policy,statistics END:VEVENT BEGIN:VEVENT UID:2023-04-19-sumana-basu@datascience.utah.edu DTSTAMP:20230419T164500Z DTSTART;TZID=America/Denver:20230419T104500 DTEND;TZID=America/Denver:20230419T114500 SUMMARY:Towards Reinforcement Learning for Precision Drug Dosing - Sumana B asu DESCRIPTION:Sumana Basu (McGill University)\n\nDrug dosing is an important application of AI\, which can be formulated as a Reinforcement Learning (R L) problem\, since every individual’s drug dosing requirement is differe nt. In this talk\, we will talk about two major challenges of using RL for drug dosing: delayed and prolonged effects of medications\, which break t he Markov assumption of the RL framework. We will talk about an approach t o solve this problem in a model free reinforcement learning setting\, talk further about the challenges of deploying it in real life and sketch the outline of a more realistic Model Based Reinforcement Learning (MBRL) solu tion to it.\n\nhttps://datascience.utah.edu/talks/2023-04-19-sumana-basu/ URL:https://datascience.utah.edu/talks/2023-04-19-sumana-basu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/92180148411?pwd=dG1OaXlZTTQ1d0M4R0RiSUpsb3k vdz09 CATEGORIES:health & medicine,machine learning END:VEVENT BEGIN:VEVENT UID:2023-05-30-bodhisattwa-majumder@datascience.utah.edu DTSTAMP:20230530T210000Z DTSTART;TZID=America/Denver:20230530T150000 DTEND;TZID=America/Denver:20230530T163000 SUMMARY:User-centric Natural Language Processing - Bodhisattwa Majumder DESCRIPTION:Bodhisattwa Majumder (UCSD)\n\nArtificial intelligence (AI) has shown remarkable effectiveness in knowledge-seeking applications (e.g.\, for recommendations and explanations). However\, the increasing expectatio n of more trust\, accessibility\, and anthropomorphism in these AI systems requires the underlying components (dialog models\, LLMs\, classifiers) t o be adaptive and adequately knowledge grounded. In reality\, the outputs of the constituent models often lack commonsense\, explanations\, and subj ectivity\, which motivates us to ask the question: what can we achieve by redesigning AI systems to start with individual needs?\n\nIdeally\, an ass istive AI system must be aware of the surrounding world\, produce faithful explanations\, and align with the user's preferences. In this talk\, I wi ll discuss a post-hoc knowledge-injection technique that enriches the dial og responses at the decoding time and promotes achieving conversational go als. Then\, I will explore how to elevate existing AI systems using a unif ied framework to map low-level and abstractive explanations by background knowledge. Finally\, I will hint at a user-centric interventionist approac h that can help users obtain more equitable predictions backed by faithful explanations as compared to a black-box counterpart. I will conclude with the future possibilities and societal impacts of next-generation user-cen tric systems.\n\nbodhi_flyer.png\n\nhttps://datascience.utah.edu/talks/202 3-05-30-bodhisattwa-majumder/ URL:https://datascience.utah.edu/talks/2023-05-30-bodhisattwa-majumder/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Where: MEB 3147 Large Conference Room Simcast: (Meeting ID: 965 3 805 0936\, Passcode: 404653) | https://utah.zoom.us/j/96538050936 CATEGORIES:fairness & ethics,natural language processing END:VEVENT BEGIN:VEVENT UID:2023-08-30-blair-sullivan@datascience.utah.edu DTSTAMP:20230830T163000Z DTSTART;TZID=America/Denver:20230830T103000 DTEND;TZID=America/Denver:20230830T114500 SUMMARY:Improving Fairness of Information Access in Networks - Blair Sulliv an DESCRIPTION:Blair Sullivan (Utah)\n\nIn social networks\, node position is a form of social capital which enables faster and more reliable access to diverse information. Structural biases often arise from network formation and can lead to significant disparities in information access based on pos ition. We discuss ways to quantify this social capital through the lens of information flow in the network\, focusing on the setting where each node may be a source of distinct desirable information. We define several meas ures of access advantage\, and consider the problem of improving equity by making interventions in the network\, focusing on the case of edge augmen tation. We describe several heuristic strategies for budgeted intervention \, and present the results of an empirical evaluation on a corpus of real- world networks.\n\nhttps://datascience.utah.edu/talks/2023-08-30-blair-sul livan/ URL:https://datascience.utah.edu/talks/2023-08-30-blair-sullivan/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:algorithms & theory,fairness & ethics,networks & graphs END:VEVENT BEGIN:VEVENT UID:2023-09-06-anna-fariha@datascience.utah.edu DTSTAMP:20230906T163000Z DTSTART;TZID=America/Denver:20230906T103000 DTEND;TZID=America/Denver:20230906T114500 SUMMARY:Blame the data\, not the system: how data constraints can help in t rustworthy machine learning and explain causes of data-system malfunction - Anna Fariha DESCRIPTION:Anna Fariha (Utah)\n\nThe core of modern data-driven systems co mprises models learned from large datasets\, and they are usually optimize d to target particular data and workloads. While these data-driven systems have seen wide adoption and success\, their reliability and proper functi on hinge on the data's continued conformance to the systems initial settin gs and assumptions. My research focuses on designing mechanisms to assess the trustworthiness of a system's inferences and explain causes of system malfunction due to data nonconformance. The key idea here is that since da ta is central to data-driven systems\, it can guide us to determine whethe r predictions made by an ML model can be trusted\, and to expose the cause of a system's unexpected behavior. In this talk\, I will talk about mecha nisms and explanation frameworks to facilitate trusting and understanding outcomes involving data and data systems.\n\nhttps://datascience.utah.edu/ talks/2023-09-06-anna-fariha/ URL:https://datascience.utah.edu/talks/2023-09-06-anna-fariha/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:machine learning END:VEVENT BEGIN:VEVENT UID:2023-09-13-david-tench@datascience.utah.edu DTSTAMP:20230913T163000Z DTSTART;TZID=America/Denver:20230913T103000 DTEND;TZID=America/Denver:20230913T114500 SUMMARY:Dynamic Graph Sketching: To Infinity And Beyond - David Tench DESCRIPTION:David Tench (Berkeley Lab)\n\nExisting graph stream processing systems must store the graph explicitly in RAM which limits the scale of g raphs they can process. The graph semi-streaming literature offers algorit hms which avoid this limitation via linear sketching data structures that use small (sublinear) space\, but these algorithms have not seen use in pr actice to date. In this talk I will explore what is needed to make graph s ketching algorithms practically useful\, and as a case study present a ske tching algorithm for connected components and a corresponding high-perform ance implementation. Finally\, I will give an overview of the many open pr oblems in this area\, focusing on improving query performance of graph ske tching algorithms.\n\nhttps://datascience.utah.edu/talks/2023-09-13-david- tench/ URL:https://datascience.utah.edu/talks/2023-09-13-david-tench/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:algorithms & theory,networks & graphs END:VEVENT BEGIN:VEVENT UID:2023-09-20-el-kindi-rezig@datascience.utah.edu DTSTAMP:20230920T163000Z DTSTART;TZID=America/Denver:20230920T103000 DTEND;TZID=America/Denver:20230920T114500 SUMMARY:Data Preparation: The Biggest Roadblock in Data Science - El Kindi Rezig DESCRIPTION:El Kindi Rezig (Utah)\n\nWhen building Machine learning (ML) mo dels\, data scientists face a significant hurdle: data preparation. ML mod els are exactly as good as the data we train them on. Unfortunately\, data preparation is tedious and laborious because it often requires human judg ment on how to proceed. In fact\, data scientists spend at least 80% of th eir time locating the datasets they want to analyze\, integrating them tog ether\, and cleaning the result.In this talk\, I will present my key contr ibutions in data preparation for data science\, which address the followin g problems: (1) data discovery: how to discover data of interest from a la rge collection of heterogeneous tables (e.g.\, data lakes)\; (2) error det ection: how to find errors in the input and intermediate data in complex d ata workflows\; and (3) data repairing: how to repair data errors with min imal human intervention. The developed systems are specifically designed t o support data science development which poses particular requirements suc h as interactivity and modularity.\n\nhttps://datascience.utah.edu/talks/2 023-09-20-el-kindi-rezig/ URL:https://datascience.utah.edu/talks/2023-09-20-el-kindi-rezig/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:data management END:VEVENT BEGIN:VEVENT UID:2023-09-27-parikshit-gopalan@datascience.utah.edu DTSTAMP:20230927T163000Z DTSTART;TZID=America/Denver:20230927T103000 DTEND;TZID=America/Denver:20230927T114500 SUMMARY:Loss Minimization and Multi-group Fairness - Parikshit Gopalan DESCRIPTION:Parikshit Gopalan (Apple Research)\n\nTraining a predictor to m inimize a loss function fixed in advance is the dominant paradigm in machi ne learning. However\, loss minimization by itself might not guarantee des iderata like fairness and accuracy that one could reasonably expect from a predictor. In contrast\, various group-fairness notions have been propsoe d that constrain the predictor to share certain statistical properties of the data\, even when conditioned on a rich family of subgroups. There is n o explicit attempt at loss minimization.In this talk\, we will explore som e recently discovered connections between loss minimization and notions of multi-group fairness. We will see settings where one can lead to the othe r\, and other settings where this is unlikely.\n\nhttps://datascience.utah .edu/talks/2023-09-27-parikshit-gopalan/ URL:https://datascience.utah.edu/talks/2023-09-27-parikshit-gopalan/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:algorithms & theory,fairness & ethics,machine learning,society & policy END:VEVENT BEGIN:VEVENT UID:2023-10-04-bryan-perozzi@datascience.utah.edu DTSTAMP:20231004T163000Z DTSTART;TZID=America/Denver:20231004T103000 DTEND;TZID=America/Denver:20231004T114500 SUMMARY:Leveraging the Structure of Data - Bryan Perozzi DESCRIPTION:Bryan Perozzi (Google Research)\n\nAlthough predictions from ma chine learning models influence more and more of our lives\, the standard way of posing a ML problem has remained relatively unchanged for decades. In the search for better models\, a new and popular family of techniques ( sometimes called Graph Machine Learning) has emerged. These techniques rel y on expanding beyond the features of an individual entity and instead loo k to pull information from its relationships. The methods offer a tantaliz ing way of improving task performance by leveraging previously unused info rmation. However\, it is not a free lunch\, as these models can be more co mplex\, difficult to train\, and may have challenges in interpretability. This talk will discuss the fundamentals of graph machine learning\, a few models\, and some insights from years of real-world applications.\n\nhttps ://datascience.utah.edu/talks/2023-10-04-bryan-perozzi/ URL:https://datascience.utah.edu/talks/2023-10-04-bryan-perozzi/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:algorithms & theory,machine learning,networks & graphs,optimizat ion END:VEVENT BEGIN:VEVENT UID:2023-10-18-mihai-budiu@datascience.utah.edu DTSTAMP:20231018T163000Z DTSTART;TZID=America/Denver:20231018T103000 DTEND;TZID=America/Denver:20231018T114500 SUMMARY:DBSP: A formal model for streaming computation and its applications to incremental computations and databases - Mihai Budiu DESCRIPTION:Mihai Budiu (Feldera)\n\nDBSP is a simple streaming programming language inspired by Digital Signal Processing [DSP]. DBSP can be used to give a precise definition of incremental computations -- operating on cha nges (deltas\, diffs). Moreover\, given a DBSP program\, a simple algorith m can convert it to a DBSP program that computes on changes. All practical database query operators (the relational algebra\, group-by\, aggregation s\, fixed-points\, etc) can be expressed in DBSP. As a consequence we obta in an algorithm which can incrementalize essentially any database query. T he DBSP theory has been formally verified using a theorem prover\, making it the first verified theory of incremental view maintenance.\n\nThe DBSP paper has received the best paper award at the 2023 conference on Very Lar ge Databases [VLDB].\n\nhttps://datascience.utah.edu/talks/2023-10-18-miha i-budiu/ URL:https://datascience.utah.edu/talks/2023-10-18-mihai-budiu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 | https://utah.zoom.us/j/91737198805pwd=Z1o2SzE4OVRodVhDW WExOTdVcUs5Zz09 CATEGORIES:algorithms & theory,data management END:VEVENT BEGIN:VEVENT UID:2023-10-25-aydin-buluc@datascience.utah.edu DTSTAMP:20231025T163000Z DTSTART;TZID=America/Denver:20231025T103000 DTEND;TZID=America/Denver:20231025T114500 SUMMARY:Computational journeys in a sparse universe - Aydin Buluc DESCRIPTION:Aydin Buluc (Berkeley Lab)\n\nSparsity is a fundamental assumpt ion that allows us to compute efficiently on and find parsimonious solutio ns to science and engineering problems. Sparsity exists in all basic scien ces such as physics\, biology\, and chemistry. I am going to give a sampli ng of recent work we have done on sparse computations. My talk will travel across diverse problem domains including randomized linear algebra\, grap h neural networks\, protein family and structure discovery from metagenomi c data\, and tensor computations. The underlying theme will be the challen ges posed by sparsity and the computational techniques we employ to overco me these challenges.\n\nhttps://datascience.utah.edu/talks/2023-10-25-aydi n-buluc/ URL:https://datascience.utah.edu/talks/2023-10-25-aydin-buluc/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 | https://utah.zoom.us/j/91737198805?pwd=Z1o2SzE4OVRodVhD WWExOTdVcUs5Zz09 CATEGORIES:biology & genomics,networks & graphs END:VEVENT BEGIN:VEVENT UID:2023-11-01-william-kuszmaul@datascience.utah.edu DTSTAMP:20231101T163000Z DTSTART;TZID=America/Denver:20231101T103000 DTEND;TZID=America/Denver:20231101T114500 SUMMARY:Linear Probing Revisited: How to Get Rid of Clustering - William Ku szmaul DESCRIPTION:William Kuszmaul (MIT)\n\nThe linear-probing hash table is one of the oldest and most widely used data structures in computer science. Ho wever\, linear probing also famously comes with a major drawback: as soon as the hash table reaches a high memory utilization\, elements within the hash table begin to cluster together\, causing insertions to become slow. This clustering phenomenon\, which was first discovered by Donald Knuth in 1962\, increases the expected time per insertion to $\\Theta(x^2)$ (rathe r than the more desirable $\\Theta(x)$) in a hash table that is a $1 - 1/x $ fraction full.\nA natural question is whether one can somehow reduce clu stering. In this talk\, we establish an even stronger statement: the class ical linear-probing hash table (even as it was first implemented in the 19 50s) already has less clustering than the classical results would seem to suggest. As insertions and deletions are performed over time\, the tombsto nes left behind by deletions cause the combinatorial structure of the hash table to stabilize in a way that eliminates clustering. This means that\, for some versions of linear probing\, the amortized expected time per ope ration is actually $\\tilde{O}(x)$. We also present a new version of linea r probing that avoids clustering entirely\, achieving $O(x)$ expected time per operation.\n\nhttps://datascience.utah.edu/talks/2023-11-01-william-k uszmaul/ URL:https://datascience.utah.edu/talks/2023-11-01-william-kuszmaul/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 | https://utah.zoom.us/j/91737198805?pwd=Z1o2SzE4OVRodVhD WWExOTdVcUs5Zz09 CATEGORIES:algorithms & theory,data management END:VEVENT BEGIN:VEVENT UID:2023-11-22-bryan-perozzi@datascience.utah.edu DTSTAMP:20231122T173000Z DTSTART;TZID=America/Denver:20231122T103000 DTEND;TZID=America/Denver:20231122T114500 SUMMARY:“Leveraging the Structure of Data“ - Bryan Perozzi DESCRIPTION:Bryan Perozzi (Google Research)\n\nAlthough predictions from ma chine learning models influence more and more of our lives\, the standard way of posing a ML problem has remained relatively unchanged for decades. In the search for better models\, a new and popular family of techniques ( sometimes called Graph Machine Learning) has emerged. These techniques rel y on expanding beyond the features of an individual entity and instead loo k to pull information from its relationships. The methods offer a tantaliz ing way of improving task performance by leveraging previously unused info rmation. However\, it is not a free lunch\, as these models can be more co mplex\, difficult to train\, and may have challenges in interpretability. This talk will discuss the fundamentals of graph machine learning\, a few models\, and some insights from years of real-world applications.\n\nhttps ://datascience.utah.edu/talks/2023-11-22-bryan-perozzi/ URL:https://datascience.utah.edu/talks/2023-11-22-bryan-perozzi/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:algorithms & theory,machine learning,networks & graphs,optimizat ion END:VEVENT BEGIN:VEVENT UID:2023-11-29-shangdi-yu@datascience.utah.edu DTSTAMP:20231129T173000Z DTSTART;TZID=America/Denver:20231129T103000 DTEND;TZID=America/Denver:20231129T114500 SUMMARY:Framework for Parallel Hierarchical Agglomerative Clustering - Shan gdi Yu DESCRIPTION:Shangdi Yu (MIT)\n\nWe study the hierarchical clustering proble m\, where the goal is to produce a dendrogram that represents clusters at varying scales of a data set. We propose the ParChain framework for design ing parallel hierarchical agglomerative clustering (HAC) algorithms\, and using the framework we obtain novel parallel algorithms for the complete l inkage\, average linkage\, and Ward's linkage criteria. Compared to most p revious parallel HAC algorithms\, which require quadratic memory\, our new algorithms require only linear memory\, and are scalable to large data se ts. ParChain is based on our parallelization of the nearest-neighbor chain algorithm\, and enables multiple clusters to be merged on every round. We introduce two key optimizations that are critical for efficiency: a range query optimization that reduces the number of distance computations requi red when finding nearest neighbors of clusters\, and a caching optimizatio n that stores a subset of previously computed distances\, which are likely to be reused. Experimentally\, we show that our highly-optimized implemen tations using 48 cores with two-way hyper-threading achieve 5.8--110.1x sp eedup over state-of-the-art parallel HAC algorithms and achieve 13.75--54. 23x self-relative speedup. Compared to state-of-the-art algorithms\, our a lgorithms require up to 237.3x less space. Our algorithms are able to scal e to data set sizes with tens of millions of points\, which previous algor ithms are not able to handle.\n\nhttps://datascience.utah.edu/talks/2023-1 1-29-shangdi-yu/ URL:https://datascience.utah.edu/talks/2023-11-29-shangdi-yu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:algorithms & theory,optimization END:VEVENT BEGIN:VEVENT UID:2023-12-06-julian-shun@datascience.utah.edu DTSTAMP:20231206T173000Z DTSTART;TZID=America/Denver:20231206T103000 DTEND;TZID=America/Denver:20231206T114500 SUMMARY:Parallel Batch-Dynamic Graph Algorithms - Julian Shun DESCRIPTION:Julian Shun (MIT)\n\nThere has been significant interest in gra ph analytics due to their applications in many domains\, including social network and Web analytics\, machine learning\, biology\, and physical simu lations. Real-world graphs today are massive and also dynamic. As many rea l-world graphs change rapidly\, it is crucial to design dynamic algorithms that efficiently maintain graph statistics upon updates\, since the cost of re-computation from scratch can be prohibitive. Furthermore\, due to th e high frequency of updates\, we can improve performance by using parallel ism to process batches of updates at a time. This talk presents new graph algorithms in this parallel batch-dynamic setting.\n\nSpecifically\, we pr esent the first parallel batch-dynamic algorithm for approximate k-core de composition that is efficient in both theory and practice. Our algorithm i s based on our novel parallel level data structure\, inspired by the seque ntial level data structures of Bhattacharya et al. and Henzinger et al. Gi ven a graph with n vertices and a batch of B updates\, our algorithm maint ains a (2 + epsilon)-approximation of the coreness values of all vertices (for any constant epsilon > 0) in O(B log^2(n)) amortized work and O(log^2 (n) loglog(n)) span (parallel time) with high probability. We implement an d experimentally evaluate our algorithm\, and demonstrate significant spee dups over state-of-the-art serial and parallel implementations for dynamic k-core decomposition.\n\nWe have also designed new parallel batch-dynamic algorithms for low out-degree orientation\, maximal matching\, clique cou nting\, graph coloring\, minimum spanning forest\, single-linkage clusteri ng\, some of which use our parallel level data structure.\n\nhttps://datas cience.utah.edu/talks/2023-12-06-julian-shun/ URL:https://datascience.utah.edu/talks/2023-12-06-julian-shun/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE CATEGORIES:algorithms & theory,networks & graphs,statistics END:VEVENT BEGIN:VEVENT UID:2023-12-13-felix-reidl@datascience.utah.edu DTSTAMP:20231213T173000Z DTSTART;TZID=America/Denver:20231213T103000 DTEND;TZID=America/Denver:20231213T114500 SUMMARY:Talk by Felix Reidl - Felix Reidl DESCRIPTION:Felix Reidl (Birkbeck University)\n\nData Science and AI have a n every increasing presence in our social\, political and economic life. G iven the immense influence the technologies of these field have and will h ave\, I argue that researchers and academic institutions should reflect on the implications their work has in the world at large.\n\nIn this talk I would like take stock of the larger context: a world lacking futures\, fra gmented academic disciplines\, and the dystopian use of technology. While we cannot hope to solve any of these problems\, I argue that we can and sh ould resist the underlying trends. To that end\, I propose that our instit utions should be constructed first and foremost around creativity and part icipation and what that could mean in practice.\n\nhttps://datascience.uta h.edu/talks/2023-12-13-felix-reidl/ URL:https://datascience.utah.edu/talks/2023-12-13-felix-reidl/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:FASB 295 CATEGORIES:society & policy END:VEVENT BEGIN:VEVENT UID:2024-01-24-anna-little@datascience.utah.edu DTSTAMP:20240124T203000Z DTSTART;TZID=America/Denver:20240124T133000 DTEND;TZID=America/Denver:20240124T143000 SUMMARY:Clustering and Visualization of High-dimensional Data using Path Me trics - Anna Little DESCRIPTION:Anna Little (Utah)\n\nThis talk will explore the utility of dat a-driven path metrics for the clustering and visualization of high-dimensi onal data. These metrics are defined by solving an optimal path problem in a proximity graph\, and are characterized by a parameter harmonizing dens ity-based and geometric features. First\, we will discuss theoretical prop erties of these metrics and implications for clustering: in particular\, s pectral clustering with path metrics leads to strong theoretical guarantee s on estimating number of clusters and cluster accuracy. Furthermore\, as the sample size converges to infinity\, the eigenvalues and eigenvectors o f the discrete path metric graph Laplacian converge to those of a continuu m operator\; this operator generates a diffusion which is accelerated in r egions of high data density\, allowing for the rapid exploration of elonga ted data structures. Secondly\, we will discuss dimension reduction with p ath metrics. Despite the allure of visually striking results from dimensio n reduction algorithms\, can we be sure the perceived patterns are genuine ly intrinsic to the data? We will see how designing algorithms which incor porate data-driven path metrics can lead to desirable and understandable p roperties for data visualization\, and specifically focus on their applica tion to the analysis of single cell RNA sequence data. Finally\, we will d iscuss generalizations of these metrics to simplex paths\, and their utili ty for multi-manifold clustering.\n\nhttps://datascience.utah.edu/talks/20 24-01-24-anna-little/ URL:https://datascience.utah.edu/talks/2024-01-24-anna-little/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/96005100565?pwd=WmFGN25RazZwV2NoMGE2dVFGMng yZz09 CATEGORIES:algorithms & theory,biology & genomics,networks & graphs,visuali zation END:VEVENT BEGIN:VEVENT UID:2024-01-31-pratik-soni@datascience.utah.edu DTSTAMP:20240131T203000Z DTSTART;TZID=America/Denver:20240131T133000 DTEND;TZID=America/Denver:20240131T143000 SUMMARY:Cryptography for Fairness - Pratik Soni DESCRIPTION:Pratik Soni (Utah)\n\nhttps://datascience.utah.edu/talks/2024-0 1-31-pratik-soni/ URL:https://datascience.utah.edu/talks/2024-01-31-pratik-soni/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/96005100565?pwd=WmFGN25RazZwV2NoMGE2dVFGMng yZz09 CATEGORIES:fairness & ethics END:VEVENT BEGIN:VEVENT UID:2024-02-07-daniel-brown@datascience.utah.edu DTSTAMP:20240207T203000Z DTSTART;TZID=America/Denver:20240207T133000 DTEND;TZID=America/Denver:20240207T143000 SUMMARY:Challenges and Progress Towards AI Alignment via Reinforcement Lear ning from Human Feedback - Daniel Brown DESCRIPTION:Daniel Brown (Utah)\n\nIn this talk I will discuss recent progr ess and challenges towards using human feedback to develop AI systems that behave in ways that are aligned with human preferences. One problem that arises when robots and other AI systems learn from human input is that the re is often a large amount of uncertainty over the human’s true intent a nd the corresponding desired AI behavior. To address this problem\, I will discuss prior and ongoing research along three main topics: (1) Enabling robots and other AI systems to learn models of human preferences in ways t hat are sample efficient and safe\, (2) Reward misidentification and causa l confusion when learning from human feedback\, and (3) Active learning me thods for learning better aligned features and reward functions.\n\nhttps: //datascience.utah.edu/talks/2024-02-07-daniel-brown/ URL:https://datascience.utah.edu/talks/2024-02-07-daniel-brown/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:GC 2560 (Gardner Commons) | https://utah.zoom.us/j/96005100565?pwd =WmFGN25RazZwV2NoMGE2dVFGMngyZz09 CATEGORIES:machine learning,robotics END:VEVENT BEGIN:VEVENT UID:2024-02-14-shivam-garg@datascience.utah.edu DTSTAMP:20240214T203000Z DTSTART;TZID=America/Denver:20240214T133000 DTEND;TZID=America/Denver:20240214T143000 SUMMARY:In-Context Learning: A Case Study of Simple Function Classes - Shiv am Garg DESCRIPTION:Shivam Garg (Harvard)\n\nIn-context learning refers to the abil ity of a model to learn new tasks from a sequence of input-output pairs gi ven in a prompt. Crucially\, this learning happens at inference time witho ut any parameter updates to the model. I will discuss our empirical effort s that shed light on some basic aspects of in-context learning: To what ex tent can Transformers\, or other models such as LSTMs be efficiently train ed to in-context learn fundamental function classes\, such as linear funct ions\, sparse linear functions\, and small decision trees? How can one eva luate in-context learning algorithms? And what are the qualitative differe nces between these architectures with respect to their ability to be train ed to perform in-context learning? This is based on joint work with Dimitr is Tsipras\, Percy Liang\, and Greg Valiant.\n\nhttps://datascience.utah.e du/talks/2024-02-14-shivam-garg/ URL:https://datascience.utah.edu/talks/2024-02-14-shivam-garg/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/96005100565?pwd=WmFGN25RazZwV2NoMGE2dVFGMng yZz09 CATEGORIES:algorithms & theory,deep learning,large language models END:VEVENT BEGIN:VEVENT UID:2024-02-21-tolga-tasdizen@datascience.utah.edu DTSTAMP:20240221T203000Z DTSTART;TZID=America/Denver:20240221T133000 DTEND;TZID=America/Denver:20240221T143000 SUMMARY:Deep Learning in Image Analysis: Applications and Challenges - Tolg a Tasdizen DESCRIPTION:Tolga Tasdizen (Utah)\n\nMachine learning and more specifically deep learning has revolutionized computer vision and images analysis prob lems in a wide range of application areas. In this talk\, we will outline four applications ranging from material science to health informatics and medical image analysis. We will discuss challenges unique to each problem and our approach using machine learning. A common theme underlying these p roblems is the difficulty of obtaining annotated data. Our contributions t owards solving this problem ranging from new semi-supervised learning form ulations to leveraging existing workflows of experts to collect annotation s will be introduced.\n\nhttps://datascience.utah.edu/talks/2024-02-21-tol ga-tasdizen/ URL:https://datascience.utah.edu/talks/2024-02-21-tolga-tasdizen/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/96005100565?pwd=WmFGN25RazZwV2NoMGE2dVFGMng yZz09 CATEGORIES:computer vision,deep learning,health & medicine,machine learning END:VEVENT BEGIN:VEVENT UID:2024-02-28-baskar-ganapathysubramanian@datascience.utah.edu DTSTAMP:20240228T203000Z DTSTART;TZID=America/Denver:20240228T133000 DTEND;TZID=America/Denver:20240228T143000 SUMMARY:Neural PDE solvers with applications in manufacturing and agricultu re - Baskar Ganapathysubramanian DESCRIPTION:Baskar Ganapathysubramanian (Iowa)\n\nNumerical simulation is a critical tool in analysis\, design\, and control of complex systems\, whi ch are usually described by partial differential equations (PDE). The trad itional approach to solving these PDEs has been via their numerical approx imations (for example\, finite difference\, finite element). Recently\, ad vances in Scientific machine learning (SciML) has opened up the possibilit y of training deep networks to solve complex PDEs. Several promising SciML approaches –for example\, PINNs\, FNO – have recently been proposed t o solve PDEs under various amounts of data availability.\n\nIn this talk\, I will discuss some of my group’s contribution to this effort in traini ng neural PDE solvers. This discussion will cover (a) formulating and trai ning a mesh-based neural network approach that solves for a large parametr ic family of PDEs\, (b) how we accelerate training a large models (for meg a voxel PDE predictions) via a method analogous to the multigrid technique used in numerical linear algebra\, (c) approaches to solve PDEs over doma ins with irregularly shaped (non-rectilinear) geometric boundaries\, and ( d) applying these approaches for solving design problems in agriculture an d energy technology. Whenever possible\, we perform analysis which reveals theoretical insights into several sources of error incurred in the model- building process.\n\nhttps://datascience.utah.edu/talks/2024-02-28-baskar- ganapathysubramanian/ URL:https://datascience.utah.edu/talks/2024-02-28-baskar-ganapathysubramani an/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/96005100565?pwd=WmFGN25RazZwV2NoMGE2dVFGMng yZz09 CATEGORIES:machine learning,optimization,physics & astronomy END:VEVENT BEGIN:VEVENT UID:2024-03-13-swabha-swayamdipta@datascience.utah.edu DTSTAMP:20240313T193000Z DTSTART;TZID=America/Denver:20240313T133000 DTEND;TZID=America/Denver:20240313T143000 SUMMARY:Understanding LLMs through their Generative Behavior\, Successes an d Shortcomings - Swabha Swayamdipta DESCRIPTION:Swabha Swayamdipta (USC)\n\nGenerative capabilities of large la nguage models have grown beyond the wildest imagination of the broader AI research community\, leading many to speculate whether these successes may be attributed to the training data or model design. I will present some w ork from my group which sheds light on understanding LLMs by studying thei r generative behavior\, successes and shortcomings. First\, I will show th at standard inference algorithms work well because of the particular desig n behind LLMs. Next\, I will discuss recently found successes and failures of LLMs on a combination of tasks\, requiring world and domain-specific k nowledge\, linguistic capabilities and awareness of human and social utili ty. Overall\, these findings paint a partial yet complex picture of our un derstanding of LLMs and provide a guide to the next steps forward.\n\nhttp s://datascience.utah.edu/talks/2024-03-13-swabha-swayamdipta/ URL:https://datascience.utah.edu/talks/2024-03-13-swabha-swayamdipta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE CATEGORIES:large language models,machine learning,natural language processi ng END:VEVENT BEGIN:VEVENT UID:2024-04-10-nitin-bakshi@datascience.utah.edu DTSTAMP:20240410T193000Z DTSTART;TZID=America/Denver:20240410T133000 DTEND;TZID=America/Denver:20240410T143000 SUMMARY:Service Operations for Justice-On-Time: A Data-Driven Queueing Appr oach - Nitin Bakshi DESCRIPTION:Nitin Bakshi (Utah)\n\nLimited resources in the judicial system can lead to costly delays\, stunted economic development\, and even failu re to deliver justice. Using the Supreme Court of India as an exemplar for such resource-constrained settings\, we apply ideas from service operatio ns to study delay. Specifically\, court dynamics constitute a case-managem ent queue\, whereby each case may experience multiple service encounters s pread across time\, but all are necessarily with the same server. Our goal is to elucidate the drivers of congestion\, focusing on metrics such as t he expected case-disposition time (delay) and expected number of cases awa iting adjudication (pendency)\, and leverage this understanding to recomme nd operational interventions.\n\nWe employ data-driven calibrated simulati ons to model the analytically intractable case-management queue. The life cycle of a case comprises two stages: pre-admission (before determining it s merit for detailed hearings) and post-admission. Our methodology allows us to capture the queueing dynamics in which the judges are shared resourc es across the two stages. It also permits modeling of holiday capacity\, w hich is flexibly tailored to address any surplus work that spills over fro m the regular year. We find that the second stage of this judicial queue i s overloaded\, but holiday capacity creates a perception of stability by s teadying performance metrics.\n\nThe sources of inefficiency that drive co ngestion include a misalignment between scheduling guidelines and judicial capacity\, coupled with the requirement to schedule hearings in advance. Together\, these factors inhibit utilization of shared capacity across the two-stage judicial queue. We demonstrate how interventions that account f or these inefficiencies can successfully tackle judicial delay. In particu lar\, scheduling to improve the allocation of time across pre- and post-ad mission cases can cut down the expected delay by as much as 65%.\n\nhttps: //datascience.utah.edu/talks/2024-04-10-nitin-bakshi/ URL:https://datascience.utah.edu/talks/2024-04-10-nitin-bakshi/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:https://utah.zoom.us/j/96005100565?pwd=WmFGN25RazZwV2NoMGE2dVFGMng yZz09 CATEGORIES:fairness & ethics,society & policy END:VEVENT BEGIN:VEVENT UID:2024-08-27-jeff-phillips@datascience.utah.edu DTSTAMP:20240827T183000Z DTSTART;TZID=America/Denver:20240827T123000 DTEND;TZID=America/Denver:20240827T133000 SUMMARY:== Sketching and Classifying Spatial Trajectories == - Jeff Phillip s DESCRIPTION:Jeff Phillips (Utah KSoC)\n\nSpatial trajectories\, often repre sented as a sequence of spatial positions\, are a standard way to represen t human mobility patterns. They also are used to represent motion patterns including for animals\, drones\, or last-mile rentals (e-scooters). Howev er\, these trajectories are notoriously difficult to work with as they ove rlap and can stretch long distances.\nIn this talk we discuss a sketch (th e minDist Sketch) that makes just about any data analysis on trajectories tasks simple and efficient. This first considers spatial trajectories as a n abstract shape\, and then maps them to a high-dimensional Euclidean spac e as a vector. We can show recovery\, pseudo-metric\, and metric propertie s of this representation. Variants can include direction information\, or traits like velocity and acceleration.\nMoreover\, once represented as thi s vector\, the trajectory data is extremely easy to work with. Allowing fo r out-of-the-box use of software for nearest-neighbor search\, clustering\ , and classification.\nIn particular\, we conduct the first formal study o f classifying spatial trajectories: given trajectories from two different distributions (e.g.\, generated by car or bus) given a new trajectory that is unlabeled\, how well can we predict which class it was from?\nOver sev eral data sets we have assembled that demand this task\, we conduct a larg e study\, and show that the minDist sketch and its variants are consistent ly the easiest and most accurate method (or at the least among the best in each instance).\n\nhttps://datascience.utah.edu/talks/2024-08-27-jeff-phi llips/ URL:https://datascience.utah.edu/talks/2024-08-27-jeff-phillips/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory END:VEVENT BEGIN:VEVENT UID:2024-09-03-esha-datta@datascience.utah.edu DTSTAMP:20240903T183000Z DTSTART;TZID=America/Denver:20240903T123000 DTEND;TZID=America/Denver:20240903T133000 SUMMARY:Topological Signatures of Out-of-Distribution Examples - Esha Datta DESCRIPTION:Esha Datta (Sandia NL)\n\nMachine learning (ML) models employed for real-world tasks will invariably encounter inference data that is dis tributionally shifted from their training datasets. Such out-of-distributi on (OOD) examples can have adverse effects on model performance and can po se significant problems in high-consequence application areas like healthc are or autonomous vehicles. We develop a topological characterization of O OD examples and present a computationally feasible methodology for detecti ng such data in a deployed pipeline. The approach leverages the known prop erty that well-trained ML models induce a topological “simplification” on its training dataset. By computing the persistent homology of the hidd en layer embeddings of training and test data\, we demonstrate empirically our ability to identify the presence of OOD examples for a given model.\n \nhttps://datascience.utah.edu/talks/2024-09-03-esha-datta/ URL:https://datascience.utah.edu/talks/2024-09-03-esha-datta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:machine learning END:VEVENT BEGIN:VEVENT UID:2024-09-10-guanhong-tao@datascience.utah.edu DTSTAMP:20240910T183000Z DTSTART;TZID=America/Denver:20240910T123000 DTEND;TZID=America/Denver:20240910T133000 SUMMARY:Are AI-enabled Systems Safe and Secure? - Guanhong Tao DESCRIPTION:Guanhong Tao\n\nAbstract\nArtificial Intelligence (AI) has been integrated into various sectors\, such as facial recognition and autonomo us driving. But are the security and safety of these AI-enabled systems fu lly ensured? In this talk\, I will present various vulnerabilities in thes e systems. My presentation will cover novel optimization techniques for id entifying and mitigating backdoor vulnerabilities in both white-box and bl ack-box settings\, achieving substantial improvements in performance. I wi ll share insights into the nature of backdoors and their presence in pre-t rained models. Finally\, I will conclude with an outlook on our recent exp loration of the security of emerging AI techniques\, such as generative AI .\n\nhttps://datascience.utah.edu/talks/2024-09-10-guanhong-tao/ URL:https://datascience.utah.edu/talks/2024-09-10-guanhong-tao/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:privacy & security END:VEVENT BEGIN:VEVENT UID:2024-09-17-aurora-clark@datascience.utah.edu DTSTAMP:20240917T183000Z DTSTART;TZID=America/Denver:20240917T123000 DTEND;TZID=America/Denver:20240917T133000 SUMMARY:The Importance of Shape in Chemistry Data - Aurora Clark DESCRIPTION:Aurora Clark\n\nData in the field of Chemistry has heavily leve raged graph theory representations within data science applications. Howev er\, there is a rich geometric and topological structure of many chemical systems (and their data) that has been less employed for feature optimizat ion\, dimensionality reduction\, and predictive models. Within this discus sion I will highlight some recent work and interests that seek to employ c omputational topology and geometry within chemistry data sets from molecul ar dynamics simulations – both in the context of ensemble average and te mporally evolving data sets.\n\nhttps://datascience.utah.edu/talks/2024-09 -17-aurora-clark/ URL:https://datascience.utah.edu/talks/2024-09-17-aurora-clark/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:networks & graphs,optimization END:VEVENT BEGIN:VEVENT UID:2024-09-24-rebecca-barter@datascience.utah.edu DTSTAMP:20240924T183000Z DTSTART;TZID=America/Denver:20240924T123000 DTEND;TZID=America/Denver:20240924T133000 SUMMARY:Veridical Data Science: the Practice of Responsible Data Analysis a nd Decision Making - Rebecca Barter DESCRIPTION:Rebecca Barter\n\nData science is often presented as a straight forward\, linear process involving statistical and computational technique s\, without addressing the complexities inherent in real-world application s. In contrast\, our new book\,\n"Veridical Data Science: The Practice of Responsible Data Analysis and Decision Making"\, teaches data scientists t o navigate the reality that most projects involve answering ambiguous doma in questions with messy data\, all while managing a complex web of human j udgment calls. We emphasize that datasets are merely approximations of rea lity\, and analyses are shaped by human interpretation. Using the Predicta bility\, Computability\, and Stability (PCS) framework to assess the trust worthiness and relevance of data-driven results\, "Veridical Data Science" provides an actionable guide for conducting responsible and trustworthy d ata science.\n\nhttps://datascience.utah.edu/talks/2024-09-24-rebecca-bart er/ URL:https://datascience.utah.edu/talks/2024-09-24-rebecca-barter/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:health & medicine,statistics END:VEVENT BEGIN:VEVENT UID:2024-10-01-simon-brewer@datascience.utah.edu DTSTAMP:20241001T183000Z DTSTART;TZID=America/Denver:20241001T123000 DTEND;TZID=America/Denver:20241001T133000 SUMMARY:Exploring long-term ecosystem change with self-organizing maps - Si mon Brewer DESCRIPTION:Simon Brewer\n\nOngoing climate change has the potential to imp act a variety of physical\, biological and social systems\, and there is i ncreasing concern that these changes may be sufficient to result in these systems crossing tipping points\, effectively undergoing irreversible chan ges in state. For slow turnover systems\, such as forest ecosystems\, unde rstanding the likelihood and ramifications of these state changes is chall enging due to the relative short observational record. Sedimentary records of ecosystem change offer an alternative data source with a wide temporal and spatial scope\, but are inherently noisy and high dimensional. Self-o rganizing maps provide a data-driven way to visualize nonlinear patterns i n these data\, and to identify past ecosystem states and state transitions . The results are used to build a simple Markov model illustrating the pro bability and directionality of these transitions.\n\nhttps://datascience.u tah.edu/talks/2024-10-01-simon-brewer/ URL:https://datascience.utah.edu/talks/2024-10-01-simon-brewer/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:biology & genomics,climate & environment,statistics END:VEVENT BEGIN:VEVENT UID:2024-10-15-raghav-venkatraman@datascience.utah.edu DTSTAMP:20241015T183000Z DTSTART;TZID=America/Denver:20241015T123000 DTEND;TZID=America/Denver:20241015T133000 SUMMARY:Minmax estimation rates for manifold learning - Raghav Venkatraman DESCRIPTION:Raghav Venkatraman\n\nThis talk is focused on obtaining minmax estimation rates for the "manifold learning" problem. Given N data points hypothesized to be i.i.d (independent and identically distributed) samples of a nice density (from within a reasonable class of densities) on a nice manifold (from within some class of nice manifolds of known intrinsic dim ension d)\, the manifold learning problem boils down to estimating certain statistics of this "ground truth" manifold\, such as the first few eigenm odes of the Laplace Beltrami operator on the manifold. The minmax estimati on problem further asks: given N such data points\, among all estimators o f the desired statistics (say\, a particular eigenvalue and associated eig enfunctions in a suitable norm)\, which one achieves the smallest maximum expected risk\, and how does this minmax risk scale in N and the intrinsic dimension d of the manifold?\n\nAn intuitive but impractical estimator co nsists in estimating the density from the given samples through a ``kernel density estimation''\, and then solving the resulting continuum eigenprob lem using a numerical method such as finite elements: this estimator turns out to be minmax optimal in scaling-- namely\, the associated expected ri sk scales like N^{-2/d+4}\, and we can show a matching lower bound for the minmax risk\, demonstrating that no estimator can do better\, in scaling\ , than this estimator.\n\nNext\, we ask: do there exist *practical* estima tors that are agnostic to knowledge of the manifold (so we don't have to d iscretize them in order to compute with finite elements!) that achieve\, a t least nearly\, this minmax scaling of the expected risk? We affirmativel y answer this question by showing that\, the spectrum of a carefully const ructed graph laplacian on a random geometric graph constructed from the gi ven N samples provides an estimator for the eigenvalue and eigenvectors of the Laplace Beltrami operator on the manifold that achieves this minmax r ate upto a log factor (to a small power). Both the lower and upper bound e stimates in the talk bring in new PDE tools to this statistical question\, and that we believe will be more broadly applicable in similar applicatio ns.\n\nThis talk is based on joint work with Nicolas Garcia Trillos and hi s PhD student Chenghui Li (U. W. Madison)\, and builds on prior joint work with Scott N. Armstrong (Courant Institute).\n\nhttps://datascience.utah. edu/talks/2024-10-15-raghav-venkatraman/ URL:https://datascience.utah.edu/talks/2024-10-15-raghav-venkatraman/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory,networks & graphs,statistics END:VEVENT BEGIN:VEVENT UID:2024-10-29-vivek-gupta@datascience.utah.edu DTSTAMP:20241029T183000Z DTSTART;TZID=America/Denver:20241029T123000 DTEND;TZID=America/Denver:20241029T133000 SUMMARY:Reasoning on Tabular and Multimodal Data - Vivek Gupta DESCRIPTION:Vivek Gupta (ASU & UCDS alumni)\n\nIn this talk\, I’ll walk t hrough some of the latest AI advancements that address the challenges of w orking with complex data\, with a focus on improving reasoning for both ta bular and multimodal data.\n\nI’ll begin by introducing H-STAR\, a hybri d algorithm that combines symbolic and semantic reasoning to enhance quest ion answering for tabular data. By leveraging multi-view table extraction and adaptive reasoning\, H-STAR has shown great potential in improving rea soning across tabular datasets.\n\nThen\, I’ll introduce MMTabQA\, a dat aset we developed to evaluate how AI systems manage multimodal tables that integrate structured text and images. Our research reveals where current models struggle to process these diverse data types\, highlighting key are as for improvement.\n\nTo wrap up\, I’ll discuss open challenges and fut ure directions\, including expanding reasoning capabilities to other compl ex data types—such as charts\, maps\, and flowcharts—and enhancing AI systems' robustness in handling numerical\, temporal\, and visual reasonin g\, particularly with large (vision) language models.\n\nhttps://datascien ce.utah.edu/talks/2024-10-29-vivek-gupta/ URL:https://datascience.utah.edu/talks/2024-10-29-vivek-gupta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:computer vision,data management,natural language processing,visu alization END:VEVENT BEGIN:VEVENT UID:2024-11-12-amir-abdullah@datascience.utah.edu DTSTAMP:20241112T193000Z DTSTART;TZID=America/Denver:20241112T123000 DTEND;TZID=America/Denver:20241112T133000 SUMMARY:Interpreting Learned Feedback Patterns in Large Language Models - A mir Abdullah DESCRIPTION:Amir Abdullah\n\nAmir is an active researcher in mechanistic in terpretability\, opening the blackbox of large language models to reverse engineer the inner workings and analyze internal representations.. In this talk\, he will discuss his paper in NeuIPS 2024 on interpreting reward mo dels in language models using sparse autoencoders on internal representati ons. Further\, Amir will introduce followup work studying whether internal representations can be transferred between large language models.\n\nhttp s://datascience.utah.edu/talks/2024-11-12-amir-abdullah/ URL:https://datascience.utah.edu/talks/2024-11-12-amir-abdullah/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:large language models,natural language processing END:VEVENT BEGIN:VEVENT UID:2024-11-19-hoaning-xue@datascience.utah.edu DTSTAMP:20241119T193000Z DTSTART;TZID=America/Denver:20241119T123000 DTEND;TZID=America/Denver:20241119T133000 SUMMARY:Computational and experimental approaches to examining short videos ' persuasive effects - Hoaning Xue DESCRIPTION:Hoaning Xue (Utah Communications)\n\nThere’s a gap in underst anding how people process multimodal information collectively\, despite ex tensive research on the effects of individual multimodal features and the rapid advances in computer vision. This gap is increasingly relevant as sh ort video platforms emerge as major information sources and influence publ ic opinion. In this talk\, I present findings from a project that combines a data-driven approach with social scientific theories to investigate how multimodal features in short videos impact audience engagement and attitu de change. This research is grounded in the theoretical framework of Messa ge Sensation Value (MSV) to theorize and quantify how multimodal features in short videos capture attention and affect information processing. This project includes two studies: (1) a computational model of MSV that predic ts video engagement from 11 multimodal features across a dataset of 15\,00 0 short videos from three popular short video platforms\; second\, an onli ne experiment examining the attentional mechanism underlying the persuasiv e effects of MSV in short videos on message credibility and attitude chang e. This project provides a useful framework and computational tool for sho rt video research.\n\nhttps://datascience.utah.edu/talks/2024-11-19-hoanin g-xue/ URL:https://datascience.utah.edu/talks/2024-11-19-hoaning-xue/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:computer vision END:VEVENT BEGIN:VEVENT UID:2024-11-26-zhichao-xu@datascience.utah.edu DTSTAMP:20241126T193000Z DTSTART;TZID=America/Denver:20241126T123000 DTEND;TZID=America/Denver:20241126T133000 SUMMARY:Representation Learning for IR and Role of Retrieval in LLM Era - Z hichao Xu DESCRIPTION:Zhichao Xu (Utah KSoC)\n\nRetrieval is the critical way of acce ssing information in people’s daily lives. Retrieval applications includ e search engines\, conversational shopping assistants\, or when asked abou t 2+3=?\, human brains do retrieval instead of reasoning. In this talk\, I ’ll briefly go through the history of representation learning in retriev al\, from Bag-of-Words representations to the latest dense and learned spa rse retrieval algorithms. Increasingly\, people go to ChatGPT or other lar ge language models for information seeking instead of search engines. With this existential crisis in mind\, I will talk about the role of retrieval in LLM era\, specifically\, retrieval-augmented generation\, strengths\, weaknesses and open problems.\n\nhttps://datascience.utah.edu/talks/2024-1 1-26-zhichao-xu/ URL:https://datascience.utah.edu/talks/2024-11-26-zhichao-xu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 1230 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:large language models,natural language processing END:VEVENT BEGIN:VEVENT UID:2025-01-17-fengjiao-wang@datascience.utah.edu DTSTAMP:20250117T203000Z DTSTART;TZID=America/Denver:20250117T133000 DTEND;TZID=America/Denver:20250117T143000 SUMMARY:Supervised Learning on Tabular Data - Fengjiao Wang DESCRIPTION:Fengjiao Wang (Utah SoC)\n\nSelf-supervised and Semi-supervised learning (SSL) on tabular data is an understudied topic. Despite some att empts\, there are two major challenges: 1. Imbalanced nature in the tabula r dataset\; 2. The one-hot encoding used in these methods becomes less eff icient for high-cardinality categorical features. To cope with the challen ges\, we propose SAWTab which uses a target encoding method\, Conditional Probability Representation (CPR)\, for efficient representation in the inp ut space of categorical features. We improve this representation by incorp orating the unlabeled samples through pseudo-labels. Furthermore\, we prop ose a Smooth Adaptive Weighting mechanism in the target encoding to mitiga te the issue of noisy and biased pseudo-labels. Experimental results on va rious datasets and comparisons with existing frameworks show that SAWTab y ields best test accuracy on all datasets. We find that pseudo-labels can h elp improve the input space representation in the SSL setting\, which enha nces the generalization of the learning algorithm.\n\nhttps://datascience. utah.edu/talks/2025-01-17-fengjiao-wang/ URL:https://datascience.utah.edu/talks/2025-01-17-fengjiao-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:machine learning END:VEVENT BEGIN:VEVENT UID:2025-01-24-dieter-fox@datascience.utah.edu DTSTAMP:20250124T210000Z DTSTART;TZID=America/Denver:20250124T140000 DTEND;TZID=America/Denver:20250124T150000 SUMMARY:Data Science & AI Day - Dieter Fox DESCRIPTION:Dieter Fox\n\nThe last years have seen astonishing progress in the capabilities of generative AI techniques\, particularly in the areas o f language and visual understanding and generation. Key to the success of these models are the use of image and text data sets of unprecedented scal e along with models that are able to digest such large datasets. We are no w seeing the first examples of leveraging such models to equip robots with open-world visual understanding and reasoning capabilities. Unfortunately \, however\, we have not achieved the RobotGPT moment\; these models still struggle with reasoning about geometry and physical interactions in the r eal world\, resulting in brittle performance on seemingly simple tasks suc h as manipulating objects in the open world. A crucial reason for this pro blem is the lack of data suitable to train powerful\, general models for r obot decision making and control. In this talk\, I will discuss approaches to generating large datasets for training robot manipulation capabilities \, with a focus on the role simulation can play in this context. I will sh ow some of our prior work\, where we demonstrated robust sim-to-real trans fer of manipulation skills trained in simulation\, and then discuss a prom ising direction toward training a model architecture that combines high-le vel\, semantic\, open-world reasoning\, with low-level 3D robot policies.\ n\nhttps://datascience.utah.edu/talks/2025-01-24-dieter-fox/ URL:https://datascience.utah.edu/talks/2025-01-24-dieter-fox/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Union Ballroom CATEGORIES:robotics END:VEVENT BEGIN:VEVENT UID:2025-02-07-omkar-bhalerao@datascience.utah.edu DTSTAMP:20250207T203000Z DTSTART;TZID=America/Denver:20250207T133000 DTEND;TZID=America/Denver:20250207T143000 SUMMARY:Triadic First-Order Logic Queries in Temporal Networks - Omkar Bhal erao DESCRIPTION:Omkar Bhalerao\n\nMotif counting is a fundamental problem in ne twork analysis\, and there is a rich literature of theoretical and applied algorithms for this problem. Given a large input network G\, a motif H is a small “pattern" graph indicative of special local structure. Motif/pa ttern mining involves finding all matches of this pattern in the input G. The simplest\, yet challenging\, case of motif counting is when H has thre e vertices\, often called a triadic query. Recent work has focused on temp oral graph mining\, where the network G has edges with timestamps (and dir ections) and H has time constraints. Such networks are common representati ons for communication networks\, citation networks\, financial transaction s\, etc.\nInspired by concepts in logic and database theory\, we introduce the study of Thresholded First Order Logic (FOL) Motif Analysis for massi ve temporal networks. A typical triadic motif query asks for the existence of three vertices that form a desired temporal pattern. An FOL motif quer y is obtained by having both existence and universal quantifiers with thre sholds. This allows for query semantics that can mine richer information f rom networks. A typical triadic query would be "find all triples of vertic es u\,v\,w such that they form a triangle within one hour". A thresholded FOL query can express "find all pairs u\,v such that for half of w where ( u\,w) formed an edge\, (v\,w) also formed an edge within an hour".\n\nWe d esign the first algorithm\, FOLTY\, for mining thresholded triadic FOL que ries\, whose theoretical running time matches the best known running time for sparse graphs. Specifically\, our algorithms run in time 𝑂 (m $\\al pha \\log \\sigma_{\\max}$). Here\, $m$ is the number of temporal edges in the input graph\, $\\alpha$ is its degeneracy (maximum core number)\, and the $\\sigma_{\\max}$ is the maximum edge multiplicity. Our procedures ca n be efficiently implemented\, and FOLTY has good empirical behavior. For example\, we can answer triadic FOL queries on graphs with nearly 70M edge s in less than an hour on commodity hardware. We believe that our work cou ld start a new research direction in the classic well-studied problem of m otif analysis.\n\nhttps://datascience.utah.edu/talks/2025-02-07-omkar-bhal erao/ URL:https://datascience.utah.edu/talks/2025-02-07-omkar-bhalerao/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory,data management,networks & graphs END:VEVENT BEGIN:VEVENT UID:2025-02-14-bao-wang@datascience.utah.edu DTSTAMP:20250214T203000Z DTSTART;TZID=America/Denver:20250214T133000 DTEND;TZID=America/Denver:20250214T143000 SUMMARY:Conditional Flow Divergence Matching - Bao Wang DESCRIPTION:Bao Wang\n\nConditional flow matching (CFM) stands out as an ef ficient simulation-free approach for training flow-based generative models \, achieving remarkable performance for data generation. However\, CFM is insufficient to ensure accuracy in learning probability paths\, and the le arned vector field significantly violates the continuity equation governin g probability flows. In response\, we establish a new total-variation boun d between the learned and ground-truth probability paths\, showing that th e gap between probability paths is bounded above by a combination of CFM l oss and an associated divergence loss. This theoretical bound informs us t o design a new objective to match both flow and divergence accompanied by an efficient implementation. Our new training approach improves the perfor mance of the flow-based generative model by a noticeable margin without si gnificantly raising the computational cost.\n\nhttps://datascience.utah.ed u/talks/2025-02-14-bao-wang/ URL:https://datascience.utah.edu/talks/2025-02-14-bao-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory,statistics END:VEVENT BEGIN:VEVENT UID:2025-02-21-chenglu-li@datascience.utah.edu DTSTAMP:20250221T203000Z DTSTART;TZID=America/Denver:20250221T133000 DTEND;TZID=America/Denver:20250221T143000 SUMMARY:Teaching and Learning at the Human-Technology Frontier: The Power o f Artificial Intelligence and Big Data - Chenglu Li DESCRIPTION:Chenglu Li (Utah Edu Psych)\n\nIn this talk\, Chenglu will exam ine the rapidly evolving field of artificial intelligence in education (AI ED) by focusing on three key research gaps: agentic AI needs\, FAccT (fair ness\, accountability\, and transparency) challenges\, and issues related to computing supremacy. He will illustrate these challenges through his cu rrent project\, ALTER-Math (AI-augmented Learning by Teaching to Enhance a nd Renovate Math Learning)\, a $10M initiative designed to accelerate midd le school math learning using generative AI-powered solutions. Additionall y\, Chenglu will discuss his contributions to advancing learning and teach ing through the development\, evaluation\, and dissemination of FAccT AI c yberinfrastructure. Finally\, he will outline promising future directions for research and collaboration in this dynamic field.\n\nhttps://datascien ce.utah.edu/talks/2025-02-21-chenglu-li/ URL:https://datascience.utah.edu/talks/2025-02-21-chenglu-li/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:education,fairness & ethics,large language models END:VEVENT BEGIN:VEVENT UID:2025-02-28-bernardo-modenesi@datascience.utah.edu DTSTAMP:20250228T203000Z DTSTART;TZID=America/Denver:20250228T133000 DTEND;TZID=America/Denver:20250228T143000 SUMMARY:Unveiling Hidden Patterns in Agent Behavior with Discrete-Choice an d Network Theory - Bernardo Modenesi DESCRIPTION:Bernardo Modenesi (UU BioStats)\n\nMany datasets in data scienc e stem from agents repeatedly making choices over time\, with each choice leading to an observable outcome. In this talk\, I introduce a novel appro ach to uncover latent agent heterogeneity\, enhancing both our understandi ng of agent behavior and causal inference estimation. By combining discret e choice models with network theory\, we develop a method to measure agent similarity based on their choice patterns. This results in a network-base d unsupervised clustering technique that groups agents with similar behavi ors—offering an interpretable alternative to black-box clustering models while maintaining explicit estimation assumptions. I will illustrate our approach using labor market data\, where workers (agents) and jobs (choice s) form a bipartite network\, with worker-job matches represented as edges . By clustering workers based on their job choices\, we can infer unobserv ed worker skills—a crucial factor in economic analysis. Through Bayesian estimation\, we reveal latent worker groups\, improving predictions of la bor market outcomes and measuring labor market discrimination more effecti vely than models relying only on observable characteristics. This seminar will detail our methodological framework\, estimation strategy\, and pract ical applications for understanding and predicting agent-choice dynamics.\ n\nBonus project:\nIn the final portion of the talk\, I will pivot to a mo re informal discussion of a preliminary 'Model Ensemble Approach to Assess ing Discrimination in Machine Learning Models' in the space of algorithmic fairness.\n\nhttps://datascience.utah.edu/talks/2025-02-28-bernardo-moden esi/ URL:https://datascience.utah.edu/talks/2025-02-28-bernardo-modenesi/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory,fairness & ethics,large language models,mach ine learning END:VEVENT BEGIN:VEVENT UID:2025-03-21-jeff-phillips@datascience.utah.edu DTSTAMP:20250321T193000Z DTSTART;TZID=America/Denver:20250321T133000 DTEND;TZID=America/Denver:20250321T143000 SUMMARY:Robust High-Dimensional Mean Estimation With Low Data Size - Jeff P hillips DESCRIPTION:Jeff Phillips (UU KSoC)\n\nRobust statistics aims to compute qu antities to represent data where a fraction of it may be arbitrarily corru pted. The most essential statistic is the mean\, and in recent years\,\nth ere has been a flurry of theoretical advancement for efficiently estimatin g the mean in high dimensions on corrupted data. While several algorithms have been proposed that\nachieve near-optimal error\, they all rely on lar ge data size requirements as a function of dimension.\n\nIn this talk\, we perform an extensive experimentation over various mean estimation techniq ues where data size might not meet this requirement due to the high-dimens ional setting.\nFor data with inliers generated from a Gaussian with known covariance\, we find experimentally that several robust mean estimation t echniques can practically improve upon the sample mean\, with the quantum entropy scaling approach from Dong et.al. (NeurIPS 2019) performing consis tently the best. However\, this consistent improvement is conditioned on a couple of simple modifications to how the steps to prune outliers work in the high-dimension\nlow-data setting\, and when the inliers deviate signi ficantly from Gaussianity. In fact\, with these modifications\, they are t ypically able to achieve roughly the same error as taking the sample mean of the uncorrupted inlier data\, even with very low data size. In addition to\ncontrolled experiments on synthetic data\, we also explore these meth ods on large language models\, deep pretrained image models\, and non-cont extual word embedding models that do not necessarily have an inherent Gaus sian distribution. Yet\, in these settings\, a mean point of a set of embe dded objects is a desirable quantity to learn\, and the data exhibits the high-dimension low-data setting studied in this paper. We show both the ch allenges of achieving this goal\, and that our updated robust mean estimat ion methods can provide\nsignificant improvement over using just the sampl e mean.\n\nhttps://datascience.utah.edu/talks/2025-03-21-jeff-phillips/ URL:https://datascience.utah.edu/talks/2025-03-21-jeff-phillips/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:natural language processing END:VEVENT BEGIN:VEVENT UID:2025-03-28-seth-pettie@datascience.utah.edu DTSTAMP:20250328T193000Z DTSTART;TZID=America/Denver:20250328T133000 DTEND;TZID=America/Denver:20250328T143000 SUMMARY:Everything you always wanted to know about Cardinality - Seth Petti e DESCRIPTION:Seth Pettie (U Michigan CS)\n\nThe Cardinality Estimation/Disti nct Elements problem is to approximate\nthe number of distinct elements in a data stream using a small\nprobabilistic data structure called a "sketc h". This problem has been\nstudied for 40 years\, has many industrial appl ications\, and is\nfeatured prominently in most courses on Big Data algori thmics. It is\ntherefore a real puzzle to explain why research on this pop ular and\nfundamental problem has been unusually slow.\n\nThis talk presen ts a complete history of the Cardinality Estimation\nproblem from Flajolet and Martin's seminal 1983 paper to the present\,\nand includes an account of how the research community became\nfractured\, delaying many natural d evelopments by decades. I will\npresent our recent efforts to achieve info rmation-theoretically\noptimal cardinality sketches\, which draws on two n otions of\n"information" developed in the 20th century: Fisher information \n(governing optimal point estimation) and Shannon entropy (governing\nopt imal space/communication).\n\nJoint work with Dingyu Wang.\n\n========\n\n https://datascience.utah.edu/talks/2025-03-28-seth-pettie/ URL:https://datascience.utah.edu/talks/2025-03-28-seth-pettie/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory,education,society & policy END:VEVENT BEGIN:VEVENT UID:2025-04-04-sabyasachi-basu@datascience.utah.edu DTSTAMP:20250404T193000Z DTSTART;TZID=America/Denver:20250404T133000 DTEND;TZID=America/Denver:20250404T143000 SUMMARY:"Triangles\, Communities\, and Dense Subgraphs" - Sabyasachi Basu DESCRIPTION:Sabyasachi Basu (UCSC)\n\nIn this talk\, we will go over a few recent results on dense subgraph discovery. We aim to discover 'many' dens e subgraphs of 'reasonable size' in real-world networks. We show that by l everaging triadic structure in graphs\, one can do this efficiently withou t complicated distributional assumptions on the input. Our techniques brid ge an important gap: most existing theory considers the setting where the number of pieces is a constant\, whereas techniques that produce `satisfac tory' decompositions rarely have density guarantees (and indeed\, often gi ve sparse\, poorly connected subgraphs). We offer a community detection fl avor to our results: we provide a new metric for the 'goodness' of communi ties in terms of density and show that the spectrum of graph matrices impl ies the existence of communities.\n\nA key goal of this talk is to unpack the several phrases in quotes in the preceding paragraph and offer some pe rspectives on why these are important (and sometimes difficult!). We also provide an algorithm that (provably) decomposes large social networks into dense subgraphs in minutes on regular laptops\, and\, time permitting\, d iscuss the setting of overlapping subgraph detection using similar techniq ues.\n\nhttps://datascience.utah.edu/talks/2025-04-04-sabyasachi-basu/ URL:https://datascience.utah.edu/talks/2025-04-04-sabyasachi-basu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:networks & graphs,society & policy END:VEVENT BEGIN:VEVENT UID:2025-04-11-peter-jacobs@datascience.utah.edu DTSTAMP:20250411T193000Z DTSTART;TZID=America/Denver:20250411T133000 DTEND;TZID=America/Denver:20250411T143000 SUMMARY:Estimating High-Dimensional Zipfians & Language Distributions - Pet er Jacobs DESCRIPTION:Peter Jacobs (UU KSoC\, Sandia)\n\nWe study estimation of large discrete distributions under the structural assumption that they follow a Zipfian distribution\, in which the ranking of alphabet items and/or leve l of decay need to be estimated from data. Empirical evidence for near Zip fian distributions has been found in diverse applications such as word and n-gram probability distributions in natural language text and chord proba bilities in musical pieces. We introduce the Sort and Snap estimator for w hen the level of decay is known but the ranking function needs to be estim ated\, and show it is minimax in several high dimensional regimes. When bo th the ranking and decay level are unknown\, we introduce an adaptive vari ant of Sort and Snap and show via Monte Carlo simulation that it outperfor ms state of the art discrete distribution estimators in these same high di mensional regimes. Our results motivate assessment of whether linguistical ly motivated marginal distributions for generating natural language that a re claimed to be Zipfian in quantitative linguistics communities are truly Zipfian. Through Monte Carlo experiments on one such well-regarded distri bution\, Sort and Snap procedures lag behind even the simplest non-paramet ric estimator (empirical proportions)\, which brings into focus that this distribution thought to be Zipfian actually departs meaningfully from the Zipfian pattern.\n\nhttps://datascience.utah.edu/talks/2025-04-11-peter-ja cobs/ URL:https://datascience.utah.edu/talks/2025-04-11-peter-jacobs/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:natural language processing,statistics END:VEVENT BEGIN:VEVENT UID:2025-04-18-tucker-hermans@datascience.utah.edu DTSTAMP:20250418T193000Z DTSTART;TZID=America/Denver:20250418T133000 DTEND;TZID=America/Denver:20250418T143000 SUMMARY:Stein Variational Inference for Robotic Learning and Control - Tuck er Hermans DESCRIPTION:Tucker Hermans (UU KSoC\, NVIDIA)\n\nProbabilistic inference\, the problem of estimating a distribution given data\, has been a central p illar of robotic algorithms for more than two decades. Inference technique s have defined the de facto standard for robotic localization\, mapping\, system calibration\, and online error estimation for mobile robots. Academ ic work has shown how these same inference problem formulations and algori thms can be used to solve problems of planning and control.\n\nIn this tal k I will discuss how probabilistic inference techniques can be extended fo r use in robotic manipulation where models may come either from engineerin g first-principles or in the form of large neural networks. I will then gi ve a brief dedication of Stein variational inference\, a recent nonparamet ric technique for probabilistic inference that is easily parallelized on m odern GPUs providing much faster inference times compared to more traditio nal Markov chain Monte Carlo methods. I will then show a few different app lications of using Stein variational inference from my lab\, including pla nning to goal distributions\, adaptive control of magnetic manipulation\, and generating diverse data for training from real-world robot failures.\n \nhttps://datascience.utah.edu/talks/2025-04-18-tucker-hermans/ URL:https://datascience.utah.edu/talks/2025-04-18-tucker-hermans/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/93909986581?pwd=d90LoHKoVAkagz1a CpJH1vTuaME9gG.1 CATEGORIES:algorithms & theory,robotics END:VEVENT BEGIN:VEVENT UID:2025-08-20-varun-shankar@datascience.utah.edu DTSTAMP:20250820T170000Z DTSTART;TZID=America/Denver:20250820T110000 DTEND;TZID=America/Denver:20250820T120000 SUMMARY:Kernel Methods for Operator Learning - Varun Shankar DESCRIPTION:Varun Shankar (Title: Kernel Methods for Operator Learning)\n\n https://datascience.utah.edu/talks/2025-08-20-varun-shankar/ URL:https://datascience.utah.edu/talks/2025-08-20-varun-shankar/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | http://utah.zoom.us/my/vsutah END:VEVENT BEGIN:VEVENT UID:2025-08-27-konstantin-genin@datascience.utah.edu DTSTAMP:20250827T170000Z DTSTART;TZID=America/Denver:20250827T110000 DTEND;TZID=America/Denver:20250827T120000 SUMMARY:Predictions as Public Reasons - Konstantin Genin DESCRIPTION:Konstantin Genin\n\nhttps://datascience.utah.edu/talks/2025-08- 27-konstantin-genin/ URL:https://datascience.utah.edu/talks/2025-08-27-konstantin-genin/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/7824755969?pwd=SUplL1EwUmU0TE5J RkhyQ2dtNmsvdz09 END:VEVENT BEGIN:VEVENT UID:2025-09-03-daniel-brown@datascience.utah.edu DTSTAMP:20250903T170000Z DTSTART;TZID=America/Denver:20250903T110000 DTEND;TZID=America/Denver:20250903T120000 SUMMARY:Swarms\, Emergent Behaviors\, and Multi-Agent Systems - Daniel Brow n DESCRIPTION:Daniel Brown\n\nhttps://datascience.utah.edu/talks/2025-09-03-d aniel-brown/ URL:https://datascience.utah.edu/talks/2025-09-03-daniel-brown/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:large language models END:VEVENT BEGIN:VEVENT UID:2025-09-10-bei-wang-phillips@datascience.utah.edu DTSTAMP:20250910T170000Z DTSTART;TZID=America/Denver:20250910T110000 DTEND;TZID=America/Denver:20250910T120000 SUMMARY:Talk by Bei Wang Phillips - Bei Wang Phillips DESCRIPTION:Bei Wang Phillips\n\nhttps://datascience.utah.edu/talks/2025-09 -10-bei-wang-phillips/ URL:https://datascience.utah.edu/talks/2025-09-10-bei-wang-phillips/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 END:VEVENT BEGIN:VEVENT UID:2025-09-17-aditya-bhaskara@datascience.utah.edu DTSTAMP:20250917T170000Z DTSTART;TZID=America/Denver:20250917T110000 DTEND;TZID=America/Denver:20250917T120000 SUMMARY:Descent with Misaligned Gradients and Applications to Hidden Convex ity - Aditya Bhaskara DESCRIPTION:Aditya Bhaskara\n\nhttps://datascience.utah.edu/talks/2025-09-1 7-aditya-bhaskara/ URL:https://datascience.utah.edu/talks/2025-09-17-aditya-bhaskara/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:optimization END:VEVENT BEGIN:VEVENT UID:2025-09-24-kenneth-blake-vernon@datascience.utah.edu DTSTAMP:20250924T170000Z DTSTART;TZID=America/Denver:20250924T110000 DTEND;TZID=America/Denver:20250924T120000 SUMMARY:Indirect dating with mixture density networks - Kenneth Blake Verno n DESCRIPTION:Kenneth Blake Vernon\n\nIt is an astonishing fact about the wor ld today that no one can say precisely how many people actually live on ou r planet\, even though it would presumably be useful to have such informat ion to plan for climate change and disaster risk management\, among other things. Luckily for us\, demographers and spatial data scientists are keen ly aware of this problem and have devised sophisticated methods for interp olating population in these regions based on their built area\, typically measured using remote sensing technology.\n\nAs it happens\, this is exact ly the reasoning applied by archaeologists seeking to reconstruct populati on sizes in the past. Unfortunately\, archaeology faces an additional chal lenge here\, since the archaeological record is a palimpsest of built area \, representing continuous human settlement over decades\, centuries\, and sometimes even millennia. So\, reconstructing population sizes across a r egion of interest requires that archaeologists also develop a chronology f or that region at the same time.\n\nA region’s chronology can be represe nted by a probability density function\, p(t)\, with well-dated archaeolog ical materials – like tree-rings and radiocarbon samples - assumed to be random draws from that distribution. Here\, we propose to estimate p usin g a deep-learning extension to the mixture model known as a Mixture Densit y Network. With this model\, we condition the chronology on diagnostic dat a X\, treating mixture parameters as unknown functions of X that can be es timated using a simple multilayer perceptron. Specifically\, we use the de nsity of X in the area around each sampled date to estimate the mixture pa rameters. This allows us to interpolate dates at under-sampled sites and t o build a better representation of the regional chronology\, one that is b ased\, in theory\, on the totality of the archaeological record.\n\nAs an example\, we fit an MDN to the distribution of tree-rings in the Mesa Verd e region of southwestern Colorado using the spatial distribution of cerami cs to estimate mixture parameters. An important ancillary goal of this res earch is to develop software tools scientists can use to train MDNs on the ir own data without also having to learn the intricacies of AI development and testing.\n\nhttps://datascience.utah.edu/talks/2025-09-24-kenneth-bla ke-vernon/ URL:https://datascience.utah.edu/talks/2025-09-24-kenneth-blake-vernon/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:geospatial END:VEVENT BEGIN:VEVENT UID:2025-10-01-luis-garcia@datascience.utah.edu DTSTAMP:20251001T170000Z DTSTART;TZID=America/Denver:20251001T110000 DTEND;TZID=America/Denver:20251001T120000 SUMMARY:A Trip to the Neural Frontier: Neurosymbolic Sensor Fusion for Trus tworthy AI-Enabled Neural Interventions - Luis Garcia DESCRIPTION:Luis Garcia\n\nhttps://datascience.utah.edu/talks/2025-10-01-lu is-garcia/ URL:https://datascience.utah.edu/talks/2025-10-01-luis-garcia/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 END:VEVENT BEGIN:VEVENT UID:2025-10-15-vineet-pandey@datascience.utah.edu DTSTAMP:20251015T170000Z DTSTART;TZID=America/Denver:20251015T110000 DTEND;TZID=America/Denver:20251015T120000 SUMMARY:Designing human-centered systems that yield data that is minimal\, relevant\, and actionable - Vineet Pandey DESCRIPTION:Vineet Pandey\n\nhttps://datascience.utah.edu/talks/2025-10-15- vineet-pandey/ URL:https://datascience.utah.edu/talks/2025-10-15-vineet-pandey/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:human-centered computing END:VEVENT BEGIN:VEVENT UID:2025-10-22-kenneth-marino@datascience.utah.edu DTSTAMP:20251022T170000Z DTSTART;TZID=America/Denver:20251022T110000 DTEND;TZID=America/Denver:20251022T120000 SUMMARY:VLM Agents - Kenneth Marino DESCRIPTION:Kenneth Marino\n\nhttps://datascience.utah.edu/talks/2025-10-22 -kenneth-marino/ URL:https://datascience.utah.edu/talks/2025-10-22-kenneth-marino/ SEQUENCE:0 STATUS:CANCELLED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:large language models END:VEVENT BEGIN:VEVENT UID:2025-10-29-anna-fariha@datascience.utah.edu DTSTAMP:20251029T170000Z DTSTART;TZID=America/Denver:20251029T110000 DTEND;TZID=America/Denver:20251029T120000 SUMMARY:Understanding Data through Change Summarization and Causal Disparit y Explanations - Anna Fariha DESCRIPTION:Anna Fariha\n\nhttps://datascience.utah.edu/talks/2025-10-29-an na-fariha/ URL:https://datascience.utah.edu/talks/2025-10-29-anna-fariha/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:causal inference,natural language processing END:VEVENT BEGIN:VEVENT UID:2025-11-05-jenny-lin@datascience.utah.edu DTSTAMP:20251105T180000Z DTSTART;TZID=America/Denver:20251105T110000 DTEND;TZID=America/Denver:20251105T120000 SUMMARY:Talk by Jenny Lin - Jenny Lin DESCRIPTION:Jenny Lin\n\nhttps://datascience.utah.edu/talks/2025-11-05-jenn y-lin/ URL:https://datascience.utah.edu/talks/2025-11-05-jenny-lin/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 END:VEVENT BEGIN:VEVENT UID:2025-11-12-kyle-dawson-tyler-hagen@datascience.utah.edu DTSTAMP:20251112T170000Z DTSTART;TZID=America/Denver:20251112T100000 DTEND;TZID=America/Denver:20251112T110000 SUMMARY:DESI: Disentangling Cosmology from Observational Artifacts - Kyle D awson\, Tyler Hagen DESCRIPTION:Kyle Dawson\, Tyler Hagen\n\nThe Dark Energy Spectroscopic Inst rument (DESI) has concluded three years of observation\, leading to the la rgest spectroscopic galaxy sample ever produced. In combination with other cosmological probes\, these measurements reveal hints of new physics beyo nd the standard cosmological model. In this talk\, we will first present t he observations and key measurements that led to these new constraints. We will then describe the role that neural networks and random forests play in this analysis and our tests of these machine learning algorithms agains t more physically-motivated\, linear models.\n\nWhere:\nThe location and t ime is moved for this talk so it can coincide with the CosmicAI seminar. I t will take place in WEB 3780 (the Evans Conference room in SCI) and at 10 am. It will also be on Zoom at this *new* link:\n\nhttps://datascience.uta h.edu/talks/2025-11-12-kyle-dawson-tyler-hagen/ URL:https://datascience.utah.edu/talks/2025-11-12-kyle-dawson-tyler-hagen/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 | https://utexas.zoom.us/j/87159746528?pwd=ouQu8lN9ARbb6a RFpvdf6Ddb1Oqa8B.1 CATEGORIES:machine learning,physics & astronomy END:VEVENT BEGIN:VEVENT UID:2025-11-19-vivek-srikumar@datascience.utah.edu DTSTAMP:20251119T180000Z DTSTART;TZID=America/Denver:20251119T110000 DTEND;TZID=America/Denver:20251119T120000 SUMMARY:Talk by Vivek Srikumar - Vivek Srikumar DESCRIPTION:Vivek Srikumar\n\nhttps://datascience.utah.edu/talks/2025-11-19 -vivek-srikumar/ URL:https://datascience.utah.edu/talks/2025-11-19-vivek-srikumar/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 END:VEVENT BEGIN:VEVENT UID:2025-12-03-madison-golden-kaylee-alexander@datascience.utah.edu DTSTAMP:20251203T180000Z DTSTART;TZID=America/Denver:20251203T110000 DTEND;TZID=America/Denver:20251203T120000 SUMMARY:Data Visualization 101 - Madison Golden\, Kaylee Alexander DESCRIPTION:Madison Golden\, Kaylee Alexander\n\nhttps://datascience.utah.e du/talks/2025-12-03-madison-golden-kaylee-alexander/ URL:https://datascience.utah.edu/talks/2025-12-03-madison-golden-kaylee-ale xander/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LNCO 1100 | https://utah.zoom.us/j/81370106930?pwd=5rAKB2C2SrgkOpr GOGuAtzeF5xxbbT.1 CATEGORIES:visualization END:VEVENT BEGIN:VEVENT UID:2026-01-09-marina-kogan@datascience.utah.edu DTSTAMP:20260109T203000Z DTSTART;TZID=America/Denver:20260109T133000 DTEND;TZID=America/Denver:20260109T143000 SUMMARY:Human-Centered Data Science for Crisis Informatics - Marina Kogan DESCRIPTION:Marina Kogan (UU KSoC\, RAI Faculty Fellow)\n\nSocial media pla tforms have been increasingly used by the public in crisis situations\, pa rtly because they upend the traditional top-down broadcasting model of ris k communication. Instead\, social media platforms facilitate a two-way inf ormation exchange between the official response channels and the general p ublic\, enabling more participatory crisis communication\, as well as coor dination and self-organization among the public. In this more complex info rmation ecosystem\, understanding the flow of information is crucial to su pporting those affected and preventing malicious actors from capitalizing on the uncertainty. However\, the study of such information flows is chall enging\, as the high-tempo\, high-volume convergent nature of crisis event s produces vast amounts of social media data\, necessitating the use of th e data science methods. On the other hand\, to glean meaningful insight fr om the crisis-related social media activity\, it is necessary to use metho ds that account for the complex social context of the user activity. In th is talk\, Kogan will show how the Human-Centered Data Science (HCDS) provi des methodological approaches that both harness the power of computational methods and account for the highly situated nature of social media activi ty in disruption. She will focus on sequence-based approaches as examples of HCDS methods in two empirical studies: analysis of attention-garnering information during a natural disaster and investigation of behavioral sign atures in coordinated information operations.\n\nhttps://datascience.utah. edu/talks/2026-01-09-marina-kogan/ URL:https://datascience.utah.edu/talks/2026-01-09-marina-kogan/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:human-centered computing,society & policy END:VEVENT BEGIN:VEVENT UID:2026-01-23-andrew-mcnutt@datascience.utah.edu DTSTAMP:20260123T203000Z DTSTART;TZID=America/Denver:20260123T133000 DTEND;TZID=America/Denver:20260123T143000 SUMMARY:Linters as Socio Technical Systems - Andrew McNutt DESCRIPTION:Andrew McNutt\n\nInterfaces—whether they are for data analysi s\, programming\, or any other activities—exist within specific communit ies of practice. The norms and standards of those groups inscribe themselv es in the form of those tools\; implicitly driving what is and is not poss ible. In this talk I will explore how linters (a spell checker-like tool u sed in programming) can be used to interrogate this arrangement. In doing so I will describe recent and on-going efforts relating to application of linters to a variety of domains\, including visualization\, color palettes \, and social media posts. Through this discussion\, I will argue that usi ng linters as a critical lens allows us to examine the technological world around us in a new light (offering new opportunities for research and des ign)\, and therein explore the values that we manifest in our tool design and selection.\n\nhttps://datascience.utah.edu/talks/2026-01-23-andrew-mcn utt/ URL:https://datascience.utah.edu/talks/2026-01-23-andrew-mcnutt/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:human-centered computing,society & policy,visualization END:VEVENT BEGIN:VEVENT UID:2026-02-06-makoto-kelp@datascience.utah.edu DTSTAMP:20260206T203000Z DTSTART;TZID=America/Denver:20260206T133000 DTEND;TZID=America/Denver:20260206T143000 SUMMARY:Navigating Advances and Inflections in Machine Learning for Atmosph eric Chemistry Modeling - Makoto Kelp DESCRIPTION:Makoto Kelp\n\nGlobal climate and Earth system models rarely in clude comprehensive atmospheric chemistry because of its high computationa l cost. A bottleneck is the chemical solver that integrates the large-dime nsional coupled systems of kinetic equations describing the chemical mecha nism. In recent years\, machine learning (ML) methods have been proposed a s a potentially transformative approach to reducing this cost by replacing traditional solvers with fast emulators. However\, early efforts showed t hat ML-based chemical solvers often suffer from rapid error growth and ins tability. In this talk\, I will review the evolving landscape of ML for at mospheric chemistry modeling over the past decade and how its trajectory m irrors broader developments in climate and weather AI. I begin with detail ing how to achieve stable emulation in 0-D box models and then show how th ese principles translate to complex global atmospheric models. I conclude by discussing the current state of ML for modeling atmospheric chemistry a nd outlining how mechanistic interpretability in geospatial foundation mod els may offer a promising future line of inquiry.\n\nhttps://datascience.u tah.edu/talks/2026-02-06-makoto-kelp/ URL:https://datascience.utah.edu/talks/2026-02-06-makoto-kelp/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:climate & environment,machine learning,optimization END:VEVENT BEGIN:VEVENT UID:2026-02-17-juliana-freire@datascience.utah.edu DTSTAMP:20260217T180000Z DTSTART;TZID=America/Denver:20260217T110000 DTEND;TZID=America/Denver:20260217T120000 SUMMARY:Dataset Discovery and Integration in the Era of Large Language Mode ls - Juliana Freire DESCRIPTION:Juliana Freire\n\n**Abstract:**\nThe proliferation of structure d data across open-data portals\, the web\, and enterprise data lakes pres ents unprecedented opportunities for scientific discovery and data-driven decision-making. However\, realizing this potential requires solving funda mental challenges in data discovery and integration: How do we find releva nt datasets among millions of candidates? How do we understand their seman tics to integrate them? And how do we build systems that are scalable\, ac curate\, and cost-effective?\n\nIn this talk\, I will present our recent w ork addressing these challenges by combining techniques from data manageme nt\, visualization\, HCI\, and modern language models. Our work is motivat ed by real-world problems across domains: supporting biomedical researcher s in integrating heterogeneous datasets\, enabling dataset search over urb an data for policy analysis and planning\, and augmenting training data to improve machine learning model performance.\n\nFirst\, I will describe ho w we are reimagining dataset discovery by designing specialized search eng ines\, developing methods to automatically derive metadata\, and introduci ng novel data-driven queries that go beyond keywords to support complex in formation needs. Then\, I will turn to data integration\, showing how we c an leverage the semantic power of Large Language Models (LLMs) to match sc hemas with state-of-the-art accuracy. I will also argue for the critical i mportance of human-AI collaboration\, demonstrating how visual analytics c an effectively place the user in the loop to guide and verify these comple x processes.\n\nFinally\, I will reflect on how LLMs are fundamentally cha nging computer science research. They make previously intractable problems like semantic data discovery and integration tractable\, yet they behave more like natural phenomena\, exhibiting variability and non-determinism. This shift forces us to move beyond deterministic algorithmic thinking and to practice computer science as a science\, adopting empirical research m ethods to understand and leverage these powerful but unpredictable tools.\ n\n**Bio:**\nJuliana Freire is an Institute Professor at the Tandon School of Engineering and Professor of Computer Science and Data Science at New York University\, where she co-directs the Visualization Imaging and Data Analysis (VIDA) Center. Her research develops methods and systems that ena ble a wide range of users to obtain trustworthy insights from data. It spa ns topics in large-scale data analysis and integration\, visualization\, m achine learning\, provenance management\, and web information discovery\, addressing application areas including urban analytics\, predictive modeli ng\, computational reproducibility\, and biomedical data harmonization. Sh e has co-authored over 250 papers\, including 12 award winners and a test- of-time award. She served as elected chair of ACM SIGMOD and as a council member of the Computing Community Consortium (CCC)\, and was the NYU lead investigator for the Moore-Sloan Data Science Environment. She is a Fellow of the ACM and AAAS\, and a winner of the ACM SIGMOD Contributions Award. Her work has been supported by funding agencies and industry partners\, i ncluding the National Science Foundation\, DARPA\, ARPA-H\, the Department of Energy\, the National Institutes of Health\, and technology companies such as Google\, Amazon\, Microsoft Research\, and IBM. Freire received he r Ph.D. and M.Sc. degrees in computer science from the State University of New York at Stony Brook and her B.S. degree in computer science from the Federal University of Ceará in Brazil.\n\nhttps://datascience.utah.edu/ta lks/2026-02-17-juliana-freire/ URL:https://datascience.utah.edu/talks/2026-02-17-juliana-freire/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 (Evans) | https://utah.zoom.us/j/85983626630 CATEGORIES:data management,large language models,natural language processin g,visualization END:VEVENT BEGIN:VEVENT UID:2026-02-19-dinesh-manocha@datascience.utah.edu DTSTAMP:20260219T223000Z DTSTART;TZID=America/Denver:20260219T153000 DTEND;TZID=America/Denver:20260219T163000 SUMMARY:Robot Navigation in the Wild - Dinesh Manocha DESCRIPTION:Dinesh Manocha\n\nIn the last few decades\, most robotics succe ss stories have been limited to structured or controlled environments. A m ajor challenge is to develop robot systems that can operate in complex or unstructured environments corresponding to homes\, dense traffic\, outdoor terrains\, public places\, etc. In this talk\, we give an overview of our ongoing work on developing robust planning and navigation technologies th at use recent advances in computer vision\, sensor technologies\, machine learning\, and motion planning algorithms. We present new methods that uti lize multi-modal observations from an RGB camera\, 3D LiDAR\, and robot od ometry for scene perception\, along with deep reinforcement learning for r eliable planning. The latter is also used to compute dynamically feasible and spatial aware velocities for a robot navigating among mobile obstacles and uneven terrains. We have integrated these methods with wheeled robots \, home robots\, and legged platforms and highlight their performance in c rowded indoor scenes\, home environments\, and dense outdoor terrains.\n\n https://datascience.utah.edu/talks/2026-02-19-dinesh-manocha/ URL:https://datascience.utah.edu/talks/2026-02-19-dinesh-manocha/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:machine learning,physics & astronomy,robotics END:VEVENT BEGIN:VEVENT UID:2026-02-24-paul-parsons@datascience.utah.edu DTSTAMP:20260224T173000Z DTSTART;TZID=America/Denver:20260224T103000 DTEND;TZID=America/Denver:20260224T113000 SUMMARY:Visualization and Judgment: Human-Centered Computing in Data-Rich P ractice - Paul Parsons DESCRIPTION:Paul Parsons\n\nVisualization and computational systems are inc reasingly powerful\, but their impact depends on a persistent\, often unde r-specified factor—human judgment. Designers frame problems and negotiat e constraints\; users interpret\, challenge\, and coordinate action around system outputs over time\, under uncertainty and constraint. In this talk \, I present a research agenda on visualization and judgment in data-rich work\, grounded in empirical studies and design-oriented analyses across t hree contexts. First\, I study data visualization design practice\, showin g how professional designers frame problems and co-evolve problem and solu tion spaces—work that is often invisible in pipeline-oriented accounts o f visualization. Second\, I examine expert judgment in complex sociotechni cal settings\, where visualization\, procedures\, and automation reshape d ecision spaces and can either support adaptive performance or encourage br ittle reliance. Third\, I extend these insights to scientific cyberinfrast ructure\, where platforms exposing advanced computation and data services succeed or fail based on whether diverse communities can understand\, adop t\, and sustain capabilities in practice. This perspective complements adv ances in modeling\, simulation\, and visual analytics by surfacing design constraints\, evaluation targets\, and failure modes that matter as system s intersect with the constraints and contingencies of practice. I close wi th directions for judgment-aware visualization and human–AI support that preserve interpretability\, accountability\, and coordination in conseque ntial domains.\n\nhttps://datascience.utah.edu/talks/2026-02-24-paul-parso ns/ URL:https://datascience.utah.edu/talks/2026-02-24-paul-parsons/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:human-centered computing,visualization END:VEVENT BEGIN:VEVENT UID:2026-02-26-zezhong-wang@datascience.utah.edu DTSTAMP:20260226T173000Z DTSTART;TZID=America/Denver:20260226T103000 DTEND;TZID=America/Denver:20260226T113000 SUMMARY:Designing Narrative-Driven Data Experiences - Zezhong Wang DESCRIPTION:Zezhong Wang\n\nIn today’s data-saturated world\, people are facing an “infodemic” of information overload. As public demand to int erpret and use data grows\, so does the urgency to rethink how we help bro ader audiences engage with data in meaningful ways. This talk explores how we can reconnect data with its storytelling roots to support understandin g and agency. Dr. Wang will present research at the intersection of visual design\, data visualization\, and human-computer interaction\, showing ho w interdisciplinary collaboration can open up new modes of communication. He will share empirical findings on how visual data narratives\, such as d ata comics\, can improve comprehension and engagement\, as well as methods for crafting visual data stories that connect data to context and lived e xperience. Examples will draw from cross-disciplinary projects in environm ental and healthcare data storytelling. The talk will conclude with future directions toward embedding narrative-driven data experiences into everyd ay tasks and decision-making.\n\nhttps://datascience.utah.edu/talks/2026-0 2-26-zezhong-wang/ URL:https://datascience.utah.edu/talks/2026-02-26-zezhong-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Evans Conference Room (WEB 3780) CATEGORIES:climate & environment,health & medicine,human-centered computing ,visualization END:VEVENT BEGIN:VEVENT UID:2026-02-27-kenny-marino@datascience.utah.edu DTSTAMP:20260227T204000Z DTSTART;TZID=America/Denver:20260227T134000 DTEND;TZID=America/Denver:20260227T144000 SUMMARY:Agents: Hype or Opportunity - Kenny Marino DESCRIPTION:Kenny Marino\n\nAre so-called "AI Agents" a fad or a potentiall y impactful research area enabled by the rapid progress in large language models? In this talk I will try to strip away the marketing copy and look at what an agent actually is\, returning to the classical understanding of the word\, and investigate how powerful new language models can present n ew opportunities for research in embodied decision making. We begin by rec ounting the places where language models have been useful in agent-like pr oblems\, then looking at the emerging environments for investigating VLM/L LM agents including computer use and robotics\, and finally discussing the frontier research challenges of LLM agents.\n\nhttps://datascience.utah.e du/talks/2026-02-27-kenny-marino/ URL:https://datascience.utah.edu/talks/2026-02-27-kenny-marino/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:large language models,natural language processing,robotics END:VEVENT BEGIN:VEVENT UID:2026-03-02-bogdan-raita@datascience.utah.edu DTSTAMP:20260302T230000Z DTSTART;TZID=America/Denver:20260302T160000 DTEND;TZID=America/Denver:20260302T170000 SUMMARY:Solving Linear PDE by Machine Learning and Commutative Algebra - Bo gdan Raita DESCRIPTION:Bogdan Raita\n\nWe use the theory of linear pde systems with co nstant coefficients (Malgrange\, Palamodov\, Pommaret\, Sturmfels) to impl ement a machine learning algorithm which generates solutions to arbitrary linear pdes. Since we preprocess the equations with computer algebra\, our methods are applicable to arbitrary pde systems\, irrespective of type (e lliptic\, hyperbolic\, etc.) or order. We test our method for classical eq uations (wave\, heat\, Laplace) and discuss future applications to equatio ns describing wave-related phenomena\, for example direct and inverse prob lems involving Maxwell’s system and the elasticity equations.\n\nht tps://datascience.utah.edu/talks/2026-03-02-bogdan-raita/ URL:https://datascience.utah.edu/talks/2026-03-02-bogdan-raita/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:LCB 222 CATEGORIES:machine learning,physics & astronomy END:VEVENT BEGIN:VEVENT UID:2026-03-03-grace-guo@datascience.utah.edu DTSTAMP:20260303T173000Z DTSTART;TZID=America/Denver:20260303T103000 DTEND;TZID=America/Denver:20260303T113000 SUMMARY:Concepts and Counterfactuals: Human-Centered Interpretability in th e Age of Foundation Models - Grace Guo DESCRIPTION:Grace Guo\n\nFoundation models are increasingly deployed in hig h-stakes domains\, yet their scale and opacity challenge traditional notio ns of AI interpretability. In this talk\, I present two complementary stra tegies for human-centered interpretability: reasoning through concepts and probing through counterfactuals. I first present MiMICRI\, a visualizatio n tool developed with doctors at Cleveland Clinic that enables them to int eractively create counterfactual medical images to examine how anatomical changes influence model predictions. By grounding explanations in domain-r elevant visual features\, this tool helps experts reason about model behav ior using their established medical knowledge. Next\, I will introduce Con cept2Concept\, a framework for auditing text-to-image models by characteri zing their outputs as distributions over named\, interpretable concepts. B y analyzing the metrics of concept frequency\, stability\, and co-occurren ce\, we uncover hidden and sometimes harmful associations in image generat ion models and real-world training datasets. Finally\, I conclude with my research agenda for developing new visualization tools and theoretical fou ndations that address the ongoing challenges of auditing and aligning the foundation models of today.\n\nhttps://datascience.utah.edu/talks/2026-03- 03-grace-guo/ URL:https://datascience.utah.edu/talks/2026-03-03-grace-guo/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Evans Conference Room (WEB 3780) CATEGORIES:causal inference,health & medicine,human-centered computing,larg e language models END:VEVENT BEGIN:VEVENT UID:2026-03-05-josh-levine@datascience.utah.edu DTSTAMP:20260305T173000Z DTSTART;TZID=America/Denver:20260305T103000 DTEND;TZID=America/Denver:20260305T113000 SUMMARY:Extracting\, Visualizing\, and Analyzing Topological Features with Discrete Representations - Josh Levine DESCRIPTION:Josh Levine\n\nTopological features provide multi-scale summari es of the behavior of continuous data from diverse applications ranging fr om astrophysics to medicine. Nevertheless\, computing them robustly is cha llenging due to numerical precision issues. A promising strategy is to fir st convert the input to a discrete representation that satisfies criteria introduced by Forman's discrete Morse theory. While numerous approaches ex ist to discretize the restricted case of gradient fields from scalar data\ , state-of-the-art algorithms for the general case of vector fields requir e expensive optimization procedures. In this talk\, I will present recent work that uses local evaluation to create discrete vector fields in linear time from two-dimensional\, triangulated vector fields. I will also frame this work within my contributions to the Topological ToolKit (TTK)\, an o pen source software platform for topological data analysis\, led by collab orators at UPMC Sorbonne.\n\nhttps://datascience.utah.edu/talks/2026-03-05 -josh-levine/ URL:https://datascience.utah.edu/talks/2026-03-05-josh-levine/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Evans Conference Room (WEB 3780) CATEGORIES:visualization END:VEVENT BEGIN:VEVENT UID:2026-03-16-tenghao-huang@datascience.utah.edu DTSTAMP:20260316T160000Z DTSTART;TZID=America/Denver:20260316T100000 DTEND;TZID=America/Denver:20260316T110000 SUMMARY:Generalizable\, Proactive\, and Agentic Learning for Open-Ended Tas ks - Tenghao Huang DESCRIPTION:Tenghao Huang\n\nOpen-ended\, human-like intelligence requires flexibility\, proactivity\, and social intelligence: we must learn subject ive goals and adapt to novel\, complex scenarios. . Current AI systems str uggle to learn these behaviors because reward signals are unclear\, task c ontext information is incomplete\, and the environment lacks observability . I outline a research agenda focused on building agentic learning systems that operate under such uncertainty. (1) Instead of enumerating task-spec ific heuristics\, I propose training AI systems through self-play in an ad versarial setting\, where models learn by interacting\, critiquing\, and i mproving against dynamically evolving counterparts. (2) I equip agents wit h the ability to proactively gather missing information when task context is incomplete and 3) I reconstruct environment representations through mem ory to support long-horizon agentic reasoning. Together\, I will show\, th ese techniques enable AI systems to learn abstract objectives in tasks as varied as creative writing\, multi-agent coordination and long-horizon pro blem solving\, enhancing the creativity\, usefulness\, and strategic helpf ulness of agents in open-ended environments.\n\nhttps://datascience.utah.e du/talks/2026-03-16-tenghao-huang/ URL:https://datascience.utah.edu/talks/2026-03-16-tenghao-huang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 (Evans) CATEGORIES:large language models END:VEVENT BEGIN:VEVENT UID:2026-03-20-xueguang-ma@datascience.utah.edu DTSTAMP:20260320T160000Z DTSTART;TZID=America/Denver:20260320T100000 DTEND;TZID=America/Denver:20260320T110000 SUMMARY:Breaking Information Silos: Advancing Search Systems for Unified In formation Seeking - Xueguang Ma DESCRIPTION:Xueguang Ma\n\nInformation seeking has been fundamental to huma n advancement\, enabling knowledge acquisition\, decision-making\, and inn ovation across disciplines. However\, traditional information retrieval sy stems often rely on specialized pipelines optimized for specific retrieval tasks\, causing information silos that hinder unified information seeking . In this talk\, I will present our work in building unified document retr ieval systems that break these information silos across three dimensions: (1) domain and language silos\, where I demonstrate how LLM-based dense re trievers achieve strong generalizability across retrieval tasks and presen t frameworks for training small\, generalizable retrievers through diverse LLM augmentation\; (2) modality silos\, where I introduce a paradigm shif t from text-based retrieval that relies on content extraction to directly encoding document screenshots\, preserving all information including text\ , images\, and layout in unified dense representations\; and (3) space sil os\, where we show the importance of LLM-powered search agents in seeking and gathering information across disparate sources\, and present fair and transparent evaluation benchmarks for assessing deep-search systems. I wil l conclude by discussing future directions that further pave the way towar d building truly unified retrieval systems for seamless information seekin g across world knowledge.\n\nhttps://datascience.utah.edu/talks/2026-03-20 -xueguang-ma/ URL:https://datascience.utah.edu/talks/2026-03-20-xueguang-ma/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MEB 3147 (LCR) CATEGORIES:large language models,natural language processing END:VEVENT BEGIN:VEVENT UID:2026-03-20-kate-isaacs@datascience.utah.edu DTSTAMP:20260320T193000Z DTSTART;TZID=America/Denver:20260320T133000 DTEND;TZID=America/Denver:20260320T143000 SUMMARY:A Matter of Audiences: Capturing and Reporting Reasoning and Result s Around Data Visualizations - Kate Isaacs DESCRIPTION:Kate Isaacs\n\nData visualizations are used throughout the data science process to facilitate the exploratory analysis and to report resu lts. Frequently\, these uses are neither separate nor solitary\, acting as a medium for data science teams to collaboratively reason about data. Thi s team-based data science work is often fast-paced and involves several fo rms of non-digital communication. I will discuss how we leverage gesture\, sketch\, and speech to create support for common meetings around data\, s pecifically collaborative remote meetings and informal presentations\, to aid this aspect of data science work. Then\, focusing on the wider dissemi nation of results\, I will discuss findings regarding the public's views o n the use of data and AI in science videos.\n\nhttps://datascience.utah.ed u/talks/2026-03-20-kate-isaacs/ URL:https://datascience.utah.edu/talks/2026-03-20-kate-isaacs/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:visualization END:VEVENT BEGIN:VEVENT UID:2026-03-23-benjie-wang@datascience.utah.edu DTSTAMP:20260323T160000Z DTSTART;TZID=America/Denver:20260323T100000 DTEND;TZID=America/Denver:20260323T110000 SUMMARY:Bridging the Formalization Gap for Generative AI - Benjie Wang DESCRIPTION:Benjie Wang\n\nGenerative models\, such as large language model s and diffusion models\, have tremendously increased the scope of problems that AI can address. As such\, there is a significant trend toward incorp orating generative AI to automate tasks across computing and more broadly\ , from controlling robotics systems\, to software generation and testing\, to searching over scientific knowledge. However\, there remains a signifi cant formalization gap between the domain knowledge\, theories\, and logic al and semantic constraints that are vital to applications\, and the stati stical patterns over natural data represented by large generative models. In this talk\, I will demonstrate how we can systematically bridge this fo rmalization gap towards more trustworthy AI. First\, drawing from examples and applications in my research\, I will show how we can utilize suitable intermediate representations of probability distributions to bridge betwe en formal language and generative models at scale. Then\, I will discuss h ow these practical methods are underpinned by my work advancing the mathem atical and computational foundations underlying these tractable representa tions of probability distributions.\n\nhttps://datascience.utah.edu/talks/ 2026-03-23-benjie-wang/ URL:https://datascience.utah.edu/talks/2026-03-23-benjie-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:large language models,natural language processing,robotics,stati stics END:VEVENT BEGIN:VEVENT UID:2026-03-25-dick-sadler@datascience.utah.edu DTSTAMP:20260325T220000Z DTSTART;TZID=America/Denver:20260325T160000 DTEND;TZID=America/Denver:20260325T170000 SUMMARY:What Happens in the 10 Years Following a State Government-Caused En vironmental Injustice? Flint's Progress Since Its Water Crisis - Dick Sadl er DESCRIPTION:Dick Sadler\n\nThe Flint Water Crisis was the result of decades of deliberate disinvestment and state government ineptitude. In this talk \, Dr. Sadler will discuss how the crisis unfolded\, and how his research - examining environmental exposures\, neighborhood conditions\, and blood lead levels - revealed its scale. He will also address what has changed in Flint since then\, including the massive new investments in the city that have brought hope in the wake of catastrophe.\n\nhttps://datascience.utah .edu/talks/2026-03-25-dick-sadler/ URL:https://datascience.utah.edu/talks/2026-03-25-dick-sadler/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:MLIB 1110 | Zoom: 890 9876 9672\, Passcode: 156565 CATEGORIES:climate & environment,society & policy END:VEVENT BEGIN:VEVENT UID:2026-03-26-bailing-lyu@datascience.utah.edu DTSTAMP:20260326T163000Z DTSTART;TZID=America/Denver:20260326T103000 DTEND;TZID=America/Denver:20260326T113000 SUMMARY:Navigating the Human-AI Nexus in Education: Bridging Cognitive Scie nce\, Learning Analytics\, and Intelligent Systems - Bailing Lyu DESCRIPTION:Bailing Lyu\n\nArtificial intelligence is increasingly transfor ming education\, reshaping how students engage with learning and how instr uctors design and deliver instruction. However\, critical gaps remain in t he field of AI in Education (AIED): (1) a frequent lack of grounding in th e cognitive and learning sciences when developing AI pedagogical tools\, ( 2) a "black box" regarding how students process and interact with AI-suppo rted environments\, and (3) a limited emphasis on meaningful human involve ment in how AI is applied in practice. This talk presents a series of rese arch projects on AI-augmented learning and teaching that address these gap s by integrating cognitive science into AI-powered educational technologie s\, examining human-AI interaction\, and centering human agency in their a pplication. Specifically\, using teachable agents as an example\, it will discuss (1) how pedagogical AI can be designed using theoretically grounde d learning principles\, (2) how students' learning processes unfold in the se environments\, and (3) how educators and learners can actively leverage AI to co-create and navigate engaging experiences that enhance learning o utcomes. Employing experimental research\, learning analytics\, and educat ional data mining\, these studies examine the intersection of human cognit ion and AI-driven learning. By aligning AI with evidence-based pedagogical strategies\, this work advances our understanding of how intelligent tech nologies can foster deeper learning\, positive learning experiences\, pers onalized instruction\, and adaptive support across diverse educational con texts.\n\nhttps://datascience.utah.edu/talks/2026-03-26-bailing-lyu/ URL:https://datascience.utah.edu/talks/2026-03-26-bailing-lyu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Evans Conference room (WEB 3780) | https://utah.zoom.us/j/87171666 093 CATEGORIES:education,human-centered computing END:VEVENT BEGIN:VEVENT UID:2026-03-27-erdogan-kaya@datascience.utah.edu DTSTAMP:20260327T163000Z DTSTART;TZID=America/Denver:20260327T103000 DTEND;TZID=America/Denver:20260327T113000 SUMMARY:From Self-Efficacy to Systemic Change: Building an Equitable Comput ing Education Research Program - Erdogan Kaya DESCRIPTION:Erdogan Kaya\n\nThis talk examines a foundational question in c omputational STEM education: to what extent can targeted interventions imp rove pre-service elementary teachers’ computational thinking teaching ef ficacy beliefs? I present a study of pre-service elementary teachers who p articipated in a three-week CT intervention integrating EV3 robotics\, Cod e.org\, and Zoombinis. Using the CTTEBI in a pre-post design\, results sho w significant gains in personal CT teaching efficacy\, pointing to importa nt directions for future work. I then present my current AI education rese arch program\, including the “Educate AI” project developing AI litera cy curriculum through linguistically inclusive elementary robotics\, the R ural AI project integrating AI concepts for rural upper elementary student s\, and the Compose with AI platform guiding grades 4-8 students in critic ally evaluating AI-generated content. I will also discuss my emerging rese arch agenda\, including several NSF proposals under review focused on AI l iteracy across K-16 settings. I close with a vision for how University of Utah’s collaborative structure across the Scientific Computing and Imagi ng (SCI) Institute\, Department of Educational Psychology\, College of Edu cation\, and partner departments represents an ideal ecosystem to advance this agenda.\n\nhttps://datascience.utah.edu/talks/2026-03-27-erdogan-kaya / URL:https://datascience.utah.edu/talks/2026-03-27-erdogan-kaya/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Evans Conference room (WEB 3780) CATEGORIES:education,fairness & ethics,robotics END:VEVENT BEGIN:VEVENT UID:2026-03-30-xiaoling-hu@datascience.utah.edu DTSTAMP:20260330T160000Z DTSTART;TZID=America/Denver:20260330T100000 DTEND;TZID=America/Denver:20260330T110000 SUMMARY:Principled Learning for Medical AI: Structure\, Reliability\, and I nterpretability - Xiaoling Hu DESCRIPTION:Xiaoling Hu\n\nThe widespread deployment of AI in medicine dema nds not only predictive accuracy but also structural awareness\, reliabili ty under uncertainty\, and interpretability for clinical trust. In this ta lk\, I will present a unified research agenda toward principled learning f or medical AI\, grounded in these core pillars.\n\nFirst\, I will discuss how incorporating explicit structure\, such as topology and spatial priors \, into neural networks enhances the model's ability to reason about fine- grained anatomical and pathological features\, which are critical for task s like brain and tumor segmentation. Second\, I will focus on reliability\ , exploring how we can quantify and mitigate uncertainty arising from impe rfect labels\, limited data\, and domain shifts\, using methods such as di stributional modeling\, hyperparameter learning\, and probabilistic infere nce. Third\, I will show how these approaches naturally support interpreta bility\, enabling AI systems to communicate meaningful representations tha t align with human clinical understanding.\n\nThrough applications in radi ology\, pathology\, neuroimaging\, and large-scale population datasets\, I will demonstrate how these principles facilitate scalable annotation\, ro bust generalization\, and scientific discovery. I will conclude with futur e directions aimed at generalizing these principles to multimodal learning \, real-world deployment\, and next-generation AI systems in medicine.\n\n https://datascience.utah.edu/talks/2026-03-30-xiaoling-hu/ URL:https://datascience.utah.edu/talks/2026-03-30-xiaoling-hu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:computer vision,health & medicine,machine learning END:VEVENT BEGIN:VEVENT UID:2026-03-31-he-yin@datascience.utah.edu DTSTAMP:20260331T164500Z DTSTART;TZID=America/Denver:20260331T104500 DTEND;TZID=America/Denver:20260331T114500 SUMMARY:Advancing GeoAI and Earth observation for environmental monitoring - He Yin DESCRIPTION:He Yin\n\nLandscapes around the world are changing rapidly\, wi th important consequences for sustainability\, climate resilience\, and so ciety. Yet monitoring these changes across regions and scales remains diff icult. In this talk\, I present a research program that combines multi-sen sor Earth observation\, geospatial artificial intelligence (GeoAI)\, and l and system science to better understand how land systems are changing\, wh at drives those changes\, and why they matter.\n\nI begin by presenting my studies using satellite image time series to map land use change and\, in collaboration with environmental scientists and ecologists\, to examine i ts implications for carbon sequestration and biodiversity. These studies a lso reveal key limitations of conventional remote sensing approaches\, inc luding sensor constraints\, limited transferability\, and scarce training data. I then show how these challenges motivate my more recent work in sen sor fusion\, physics-informed machine learning\, and deep learning with ve ry-high-resolution imagery — applied to problems ranging from irrigation water use and wildfire-invasive species interactions to conflict-induced environmental damage. I conclude by discussing the broader goal of buildin g GeoAI models for environmental monitoring that are informed by physical processes\, transferable across contexts\, and useful for real-world decis ion-making.\n\nhttps://datascience.utah.edu/talks/2026-03-31-he-yin/ URL:https://datascience.utah.edu/talks/2026-03-31-he-yin/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Evans Conference room (WEB 3780) CATEGORIES:climate & environment,geospatial,machine learning,society & poli cy END:VEVENT BEGIN:VEVENT UID:2026-04-01-md-mostafijur-rahman@datascience.utah.edu DTSTAMP:20260401T163000Z DTSTART;TZID=America/Denver:20260401T103000 DTEND;TZID=America/Denver:20260401T113000 SUMMARY:Efficient and Reliable AI for Real-World Healthcare Deployment - Md Mostafijur Rahman DESCRIPTION:Md Mostafijur Rahman\n\nHealthcare is one of the highest-impact domains for AI\, yet reliable deployment at scale remains difficult. To t ruly improve patient care and clinical workflows\, AI must operate under r eal clinical constraints\, not just in ideal lab settings. In practice\, d eployment is limited by high compute and memory costs\, scarce labeled dat a\, and distribution shifts across sites and time. Many clinically importa nt findings are also rare and long-tailed\, which makes generalization esp ecially challenging. My research makes deployability a design objective by developing methods that stay accurate under strict resource and data cons traints. In this talk\, I will first discuss high-performance lightweight deep learning architectures built by redesigning core building blocks. I w ill then present training-time generative supervision strategies that impr ove data efficiency and generalization to rare and long-tailed cases with no inference overhead. I will conclude with a forward-looking direction to ward real-time perception for surgical assistance\, where reliable perform ance under strict constraints is non-negotiable.\n\nhttps://datascience.ut ah.edu/talks/2026-04-01-md-mostafijur-rahman/ URL:https://datascience.utah.edu/talks/2026-04-01-md-mostafijur-rahman/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:health & medicine,machine learning END:VEVENT BEGIN:VEVENT UID:2026-04-06-qiang-ji@datascience.utah.edu DTSTAMP:20260406T163000Z DTSTART;TZID=America/Denver:20260406T103000 DTEND;TZID=America/Denver:20260406T113000 SUMMARY:Towards Data-Efficient\, Trustworthy\, and Generalizable AI for Vis ual Understanding - Qiang Ji DESCRIPTION:Qiang Ji\n\nArtificial Intelligence (AI) has achieved remarkabl e progress and is increasingly integrated across a wide range of fields\, fueling what many describe as the fourth industrial revolution. However\, behind this widespread enthusiasm lie fundamental limitations. Today’s A I systems face three major challenges: (1) an insatiable demand for large- scale labeled data\, (2) limited trustworthiness due to inadequate uncerta inty quantification\, and (3) poor generalization across domains. These ch allenges cannot be addressed simply by scaling data and computation\; inst ead\, they require foundational advances in theory and methodology.\n\nIn this talk\, I will present recent research from my lab that addresses thes e challenges in a variety of computer vision tasks. To improve data effici ency and generalization\, I will introduce our work on knowledge-augmented deep learning\, where prior knowledge from diverse sources is systematica lly identified\, encoded\, and integrated with data-driven neural networks . This approach leads to hybrid neural-symbolic models that are both more data-efficient and more generalizable. To enhance model trustworthiness an d explainability\, I will discuss our advances in Bayesian deep learning. First\, I will present our work on a Bayesian Transformer framework for ac curate and robust human activity recognition. I will then introduce our wo rk on uncertainty attribution\, which identifies the sources of uncertaint y in deep models and leverages this information for uncertainty mitigation and improved model performance. Finally\, I will highlight our recent wor k on causal deep learning for addressing domain generalization. I will int roduce a neural causal model that learns domain-invariant representations by identifying and eliminating spurious correlations arising from data bia ses.\n\nTogether\, these efforts aim to advance a new generation of AI sys tems that are more data-efficient\, trustworthy\, and robust\, enabling re liable deployment across a range of domains including human behavior under standing\, medical imaging\, scientific discovery\, and human–robot inte raction.\n\nhttps://datascience.utah.edu/talks/2026-04-06-qiang-ji/ URL:https://datascience.utah.edu/talks/2026-04-06-qiang-ji/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:causal inference,computer vision,deep learning,machine learning END:VEVENT BEGIN:VEVENT UID:2026-04-07-si-chen@datascience.utah.edu DTSTAMP:20260407T163000Z DTSTART;TZID=America/Denver:20260407T103000 DTEND;TZID=America/Denver:20260407T113000 SUMMARY:Advancing AI Literacy and Human-Centered AI for Teaching and Learni ng - Si Chen DESCRIPTION:Si Chen\n\nArtificial intelligence (AI) is rapidly transforming education and how people learn\, teach\, and prepare for the future. Yet many systems are still built around technical capabilities rather than the real needs of students\, educators\, and families. In this talk\, I prese nt a human-centered design research agenda that advances both AI literacy and AI for teaching and learning across students\, families\, and educator s. First\, I define and conceptualize generative AI literacy across childr en and parents by co-designing measurement frameworks and interactive tool s. This work enables families to build a shared understanding of AI while supporting its critical and responsible use in everyday self-directed lear ning contexts. Second\, I present AI Academy\, an faculty-facing professio nal development program and badge system that expands institutional capaci ty for AI. Through curriculum design and lightweight tools\, this work sup ports faculty in integrating generative AI into teaching while strengtheni ng their AI literacy and maintaining pedagogical and disciplinary goals. L astly\, I present AI-powered tutoring systems and other AI interactions I have developed with college learners with disabilities\, including LLM-bas ed chatbots\, particularly for Deaf and Hard of Hearing students.The talk is intended for a broad\, interdisciplinary audience.\n\nhttps://datascien ce.utah.edu/talks/2026-04-07-si-chen/ URL:https://datascience.utah.edu/talks/2026-04-07-si-chen/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:education,human-centered computing,large language models END:VEVENT BEGIN:VEVENT UID:2026-04-08-fahim-faisal@datascience.utah.edu DTSTAMP:20260408T160000Z DTSTART;TZID=America/Denver:20260408T100000 DTEND;TZID=America/Denver:20260408T110000 SUMMARY:Multilingual Model Adaptation for Under-Served Languages - Fahim Fa isal DESCRIPTION:Fahim Faisal\n\nLanguage models with multilingual capabilities serve as crucial touchpoints for improving the inclusion of underrepresent ed languages in Natural Language Processing (NLP). This research investiga tes the structural sources of linguistic underrepresentation and explores strategies for improving the adaptation of low-resource language varieties through the development of linguistically grounded resources. We first ex amine the extent of multilingual and geographic representation gaps across three key dimensions of language modeling: datasets\, model architecture\ , and model-generated text. Next\, we introduce DialectBench\, an initiati ve designed to evaluate language variation in the form of dialects and lan guage varieties—an aspect often overlooked in NLP benchmarks\, which pri marily focus on standardized language forms. To address the challenges rev ealed by these analyses\, we propose two adaptation frameworks. The first introduces phylogenetic adapter hierarchies that exploit language-family s tructure to enable zero-shot transfer across related languages. The second presents a pivot-based reinforcement learning approach that leverages hig h-resource expert models to transfer reasoning alignment without requiring target-language annotations. Together\, these linguistically motivated ad aptation strategies aim to improve the performance of language models on u nderrepresented languages and dialects\, ultimately contributing to more e quitable and accessible NLP systems for diverse linguistic communities.\n\ nhttps://datascience.utah.edu/talks/2026-04-08-fahim-faisal/ URL:https://datascience.utah.edu/talks/2026-04-08-fahim-faisal/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:fairness & ethics,large language models,natural language process ing,society & policy END:VEVENT BEGIN:VEVENT UID:2026-04-10-chase-neumann@datascience.utah.edu DTSTAMP:20260410T193000Z DTSTART;TZID=America/Denver:20260410T133000 DTEND;TZID=America/Denver:20260410T143000 SUMMARY:Building and Utilizing Foundation Models for Drug Discovery and Cli nical Development - Chase Neumann DESCRIPTION:Chase Neumann (PhD)\n\nThe journey toward decoding biology at s cale began with a focus on high-dimensional cellular morphology. At Recurs ion\, our differentiation centered on deep learning models designed to lea rn biological representations directly from imaging\, enabling predictive inference at a massive scale. By leading multiple cross-functional teams f rom early-stage discovery through to early clinical development\, we demon strated the power of this "inference-first" philosophy. A primary highligh t of these efforts was the RBM39 program\, a novel molecular glue degrader discovered entirely through computational inference rather than tradition al screening. This success proved that models could identify complex biolo gical mechanisms\; however\, moving from cellular discovery to comprehensi ve patient care requires a leap into even higher-dimensional\, clinical da ta.\n\nValinor represents the next evolution of this mission. We build mul timodal clinical foundation models\, co-designing data collection and mode l architecture to maximize signal within and across complex modalities. We have proprietary access to patient cohorts and collaborate with biobanks\ , clinical trial sites\, and academic partners\, giving us unique data adv antages at scale. Our approach is to first build the best unimodal patient representations across modalities—including DNA\, transcriptomics\, pro teomics\, cfDNA\, histopathology\, and patient reports—and then fuse the m.\n\nWe have shown that attention-based fusion consistently outperforms u nimodal approaches while making the contributions of different modalities interpretable. This enables genuine clinical reasoning: the model can chai n evidence across modalities\, explain which features drive a prediction\, and engage with clinicians in natural language. Ultimately\, we envision a virtual patient that reasons over the full spectrum of a patient's biolo gy the way an expert clinician would\, but at a scale and resolution no hu man can match.\n\nhttps://datascience.utah.edu/talks/2026-04-10-chase-neum ann/ URL:https://datascience.utah.edu/talks/2026-04-10-chase-neumann/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:biology & genomics,health & medicine,large language models,natur al language processing END:VEVENT BEGIN:VEVENT UID:2026-04-13-jihyun-rho@datascience.utah.edu DTSTAMP:20260413T190000Z DTSTART;TZID=America/Denver:20260413T130000 DTEND;TZID=America/Denver:20260413T140000 SUMMARY:Supporting responsible use of AI in visual-based teaching and learn ing - Jihyun Rho DESCRIPTION:Jihyun Rho\n\nArtificial intelligence (AI) is increasingly inte grated into educational contexts\, reshaping how teachers and students gen erate and interpret visual representations such as diagrams\, illustration s\, and data visualizations. While AI offers efficiency and flexibility\, it also introduces inaccuracies and misleading interpretations that can un dermine learning. In this talk\, I present a research program centered on supporting the responsible use of AI in visual-based teaching and learning by promoting visual literacy. Grounded in design-based research\, I desig n and evaluate AI-augmented learning environment where educators and stude nts engage with AI-generated outputs as critical inquirers. This work show s that these approaches improved interpretive accuracy\, deepened reasonin g about visual representations\, and reduce over-reliance on AI. Together\ , this work contributes design principles for fostering responsible AI use in visual-based educational contexts.\n\nhttps://datascience.utah.edu/tal ks/2026-04-13-jihyun-rho/ URL:https://datascience.utah.edu/talks/2026-04-13-jihyun-rho/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:education,fairness & ethics END:VEVENT BEGIN:VEVENT UID:2026-04-14-sameer-honwad@datascience.utah.edu DTSTAMP:20260414T163000Z DTSTART;TZID=America/Denver:20260414T103000 DTEND;TZID=America/Denver:20260414T113000 SUMMARY:Move Slow and Build Community - Sameer Honwad DESCRIPTION:Sameer Honwad\n\nSocio-emotional learning (SEL) and reflection are foundational components of quality education\, yet they remain underre presented in many school curricula. Reflection supports scientific thinkin g and inquiry\, while SEL equips students with the emotional resilience ne eded to take risks\, embrace failure\, and navigate the uncertainties inhe rent in learning and innovation. Together\, these competencies are essenti al not only for academic and professional growth\, but for everyday wellbe ing.Implementing SEL and reflective practices in rural India presents dist inct challenges. Limited access to trained professionals and consistent re sources is compounded by cultural stigma around openly discussing emotions \, barriers that make it difficult for students to develop these critical skills in traditional school settings.\n\nTo address this gap\, we designe d a conversational chatbot tailored for students in rural India\, providin g a low-barrier\, accessible space for daily reflection and emotional expr ession. Importantly\, the chatbot is also designed to serve as a springboa rd for teachers to introduce concepts of AI literacy\, enabling meaningful classroom conversations about how AI systems work\, their limitations\, a nd the ethical considerations surrounding their use. This presentation pre sents the co-design process behind the chatbot\, highlighting the collabor ative contributions of an interdisciplinary team comprising teachers\, the rapists\, computer scientists\, and learning scientists. The presentation discusses how this cross-disciplinary approach shaped a tool intended to h elp students engage with their socio-emotional selves and build a habit of reflective thinking in their daily lives.\n\nhttps://datascience.utah.edu /talks/2026-04-14-sameer-honwad/ URL:https://datascience.utah.edu/talks/2026-04-14-sameer-honwad/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:education,society & policy END:VEVENT BEGIN:VEVENT UID:2026-04-15-shiqi-yu@datascience.utah.edu DTSTAMP:20260415T160000Z DTSTART;TZID=America/Denver:20260415T100000 DTEND;TZID=America/Denver:20260415T103000 SUMMARY:AI for Multi-Wavelength X-ray Analysis - Shiqi Yu DESCRIPTION:Shiqi Yu\n\nAccurate parameter estimation in X-ray astronomy ty pically relies on traditional methods\, such as likelihood-based spectral fitting\, which can be computationally prohibitive as model complexity and data dimensionality increase. In this talk\, I present a neural network-b ased framework designed to bypass iterative fitting by directly mapping sp ectral observations to physical parameters. Using the Circinus galaxy as a benchmark\, we demonstrate how architectures trained on synthetic data fr om theoretical models can recover intrinsic properties\, such as column de nsity and torus geometry\, with both high speed and high precision. This a pproach maintains the physical rigor required for broadband analysis with multiple telescopes while significantly reducing inference time. I will di scuss the challenges and resolutions associated with training and predicti ng on multi-instrument data\, as well as the potential for these AI-driven methods to enable large-scale systematic studies across various astrophys ical sources and other scientific applications.\n\nhttps://datascience.uta h.edu/talks/2026-04-15-shiqi-yu/ URL:https://datascience.utah.edu/talks/2026-04-15-shiqi-yu/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Dinosaur Room (CSC 206) | https://utexas.zoom.us/j/84742203545?pwd =lUosaf3T6bkIS1QAaIQYHYiiClE2ZP.1 CATEGORIES:deep learning,machine learning,physics & astronomy END:VEVENT BEGIN:VEVENT UID:2026-04-15-amirali-abdullah@datascience.utah.edu DTSTAMP:20260415T163000Z DTSTART;TZID=America/Denver:20260415T103000 DTEND;TZID=America/Denver:20260415T110000 SUMMARY:Controlling LLM's via Activation Geometry - Amirali Abdullah DESCRIPTION:Amirali Abdullah\n\nControlling the behavior of large language models at inference time is an increasingly important problem. In this tal k\, I present a simple and unified approach to steering model behavior bas ed on activation geometry. By learning a single classifier over hidden rep resentations\, we can derive directions that control multiple attributes s uch as helpfulness\, style\, or safety\, and compose them dynamically with out retraining.\nThis framework enables flexible\, low cost control of mod el outputs and highlights a geometric view of representation space beyond fixed linear directions. I will discuss empirical results showing how this approach supports multi attribute control in practice\, and briefly outli ne how such steering mechanisms can be useful in scientific settings where reliable and interpretable model behavior is critical. Our recent followu p work suggests that similar activation level interventions can extend acr oss modalities\, enabling systematic analysis and control in text to image models through composable operations.\n\nhttps://datascience.utah.edu/tal ks/2026-04-15-amirali-abdullah/ URL:https://datascience.utah.edu/talks/2026-04-15-amirali-abdullah/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:Dinosaur Room (CSC 206) | https://utexas.zoom.us/j/84742203545?pwd =lUosaf3T6bkIS1QAaIQYHYiiClE2ZP.1 CATEGORIES:large language models,natural language processing END:VEVENT BEGIN:VEVENT UID:2026-04-15-chengbin-deng@datascience.utah.edu DTSTAMP:20260415T193000Z DTSTART;TZID=America/Denver:20260415T133000 DTEND;TZID=America/Denver:20260415T143000 SUMMARY:AI-Driven Environmental Intelligence: Scalable and System-Level App roaches for Environmental Decision Making - Chengbin Deng DESCRIPTION:Chengbin Deng\n\nEnvironmental systems are becoming more comple x\, dynamic\, and tightly connected to human activities\, yet much of our current work remains focused on isolated models or static mapping. In this talk\, a framework will be present and discussed that uses AI to move bey ond observation toward a more integrated understanding of environmental sy stems. The central idea is to link geospatial data\, models\, and real-wor ld decisions so that environmental information can better reflect changing conditions across space and time while remaining meaningful for researche rs and stakeholders. By drawing on some recent federally supported project s\, this talk will show how this framework can capture large scale environ mental dynamics\, account for social and local context\, and support more informed decision processes in various settings\, ranging from urban syste ms to extreme events. Instead of using AI just as a tool\, this work views it as part of a broader system that shapes how environmental problems are understood and acted\, with the goal of advancing a more adaptive and pra ctical form of environmental intelligence.\n\nhttps://datascience.utah.edu /talks/2026-04-15-chengbin-deng/ URL:https://datascience.utah.edu/talks/2026-04-15-chengbin-deng/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 CATEGORIES:climate & environment,geospatial END:VEVENT BEGIN:VEVENT UID:2026-04-17-daniel-sieta@datascience.utah.edu DTSTAMP:20260417T193000Z DTSTART;TZID=America/Denver:20260417T133000 DTEND;TZID=America/Denver:20260417T143000 SUMMARY:Multimodal Data Augmentation for Data-Efficient Robot Manipulation - Daniel Sieta DESCRIPTION:Daniel Sieta (USC)\n\nDespite recent advances\, learning-based robot manipulation systems often require large demonstration datasets and degrade in cluttered or deformable environments. This talk presents diffus ion-based multimodal data augmentation methods that synthesize consistent observations and action labels. By augmenting limited demonstrations\, the se approaches substantially reduce data requirements and enable robust man ipulation in complex\, real-world settings.\n\nhttps://datascience.utah.ed u/talks/2026-04-17-daniel-sieta/ URL:https://datascience.utah.edu/talks/2026-04-17-daniel-sieta/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB L112 | https://utah.zoom.us/j/85983626630 CATEGORIES:computer vision,robotics END:VEVENT BEGIN:VEVENT UID:2026-08-28-warren-pettine@datascience.utah.edu DTSTAMP:20260828T193000Z DTSTART;TZID=America/Denver:20260828T133000 DTEND;TZID=America/Denver:20260828T143000 SUMMARY:About Start-UP MTN: which involves Comptutational Neuroscience and AI - Warren Pettine DESCRIPTION:Warren Pettine (UU Psychiatry & MTN)\n\nhttps://datascience.uta h.edu/talks/2026-08-28-warren-pettine/ URL:https://datascience.utah.edu/talks/2026-08-28-warren-pettine/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 CATEGORIES:biology & genomics END:VEVENT BEGIN:VEVENT UID:2026-09-04-george-vega-yon@datascience.utah.edu DTSTAMP:20260904T193000Z DTSTART;TZID=America/Denver:20260904T133000 DTEND;TZID=America/Denver:20260904T143000 SUMMARY:Data Science of Tracking Measles in Utah - George Vega Yon DESCRIPTION:George Vega Yon (UU Epidemiology)\n\nFrom high-performance comp uting clusters to the constraints of a local health department\, this talk presents a research program in statistical computing and data science\, s panning network science and public health\, and goes deep on one example f rom each. On the public health side\, agent-based simulation developed wit h the Utah Department of Health and Human Services turns vaccination and d emographic data into school-level measles risk\, routed through the state' s health districts to the school districts that act on it\; related outbre ak-reconstruction work along the Utah-Nevada border estimated infections m issed by surveillance\, producing results consistent with independent geno mic analyses. On the network science side\, Exponential-family Random Grap h Models (ERGMs)—statistical models used to characterize the structure o f observed networks—I present a novel application of a pooled bipartite model of patient-provider networks in healthcare\, highlighting methodolog ical challenges in data heterogeneity and model convergence. Underneath bo th is scientific software built for other researchers to use: epiworld\, e rgmito\, netdiffuseR\, rgexf\, slurmR\, and others\, under a single commit ment—the same model should run on a high-performance cluster and inside a health department with far less to work with. Making powerful methods ex ecutable in constrained environments is treated here not as an implementat ion detail\, but as a methodological contribution in its own right.\n\nhtt ps://datascience.utah.edu/talks/2026-09-04-george-vega-yon/ URL:https://datascience.utah.edu/talks/2026-09-04-george-vega-yon/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 CATEGORIES:health & medicine,networks & graphs,society & policy,statistics END:VEVENT BEGIN:VEVENT UID:2026-09-11-benjie-wang@datascience.utah.edu DTSTAMP:20260911T193000Z DTSTART;TZID=America/Denver:20260911T133000 DTEND;TZID=America/Denver:20260911T143000 SUMMARY:The Formalization Gap in Generative AI - Benjie Wang DESCRIPTION:Benjie Wang (UU Computing)\n\nGenerative models\, such as large language models and diffusion models\, have tremendously increased the sc ope of problems that AI can address. As such\, there is a significant tren d toward incorporating generative AI to automate tasks across computing an d more broadly\, from controlling robotics systems\, to software generatio n and testing\, to searching over scientific knowledge. However\, there re mains a significant **formalization gap** between the domain knowledge\, t heories\, and logical and semantic constraints that are vital to applicati ons\, and the statistical patterns over natural data represented by large generative models. In this talk\, I will demonstrate how we can systematic ally bridge this formalization gap towards more reliable and efficient AI. First\, drawing from examples and applications in my research\, I will sh ow how to design practical inference-time strategies using probabilistic r easoning that bridge between formal language and generative models at scal e. Then\, I will discuss how these practical methods are underpinned by ad vances in the mathematical and computational foundations of efficiently re presenting joint probability distributions.\n\nhttps://datascience.utah.ed u/talks/2026-09-11-benjie-wang/ URL:https://datascience.utah.edu/talks/2026-09-11-benjie-wang/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 CATEGORIES:deep learning,large language models,natural language processing, robotics END:VEVENT BEGIN:VEVENT UID:2026-09-18-polina-kukhareva@datascience.utah.edu DTSTAMP:20260918T193000Z DTSTART;TZID=America/Denver:20260918T133000 DTEND;TZID=America/Denver:20260918T143000 SUMMARY:Using AI and Real-World Data to Personalize Diabetes Treatment - Po lina Kukhareva DESCRIPTION:Polina Kukhareva (UU BMI)\n\nPatients with type 2 diabetes have increasingly many treatment options\, but responses to these therapies va ry substantially across individuals and across outcomes such as glycemic c ontrol\, weight\, cardiovascular events\, kidney outcomes\, and hypoglycem ia. In this talk\, I will describe our work using large-scale electronic h ealth record data\, causal inference\, and machine learning to characteriz e treatment heterogeneity and develop approaches for individualized treatm ent selection. I will discuss challenges in reconstructing longitudinal tr eatment regimens from real-world prescribing data\, predicting multiple pa tient outcomes\, and identifying the level of patient-specific resolution that can be reliably supported by available data. I will also discuss how these methods can ultimately support patient-centered treatment comparison s and clinical decision making.\n\nhttps://datascience.utah.edu/talks/2026 -09-18-polina-kukhareva/ URL:https://datascience.utah.edu/talks/2026-09-18-polina-kukhareva/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 CATEGORIES:causal inference,health & medicine,machine learning END:VEVENT BEGIN:VEVENT UID:2026-09-25-matthew-s-sigman@datascience.utah.edu DTSTAMP:20260925T193000Z DTSTART;TZID=America/Denver:20260925T133000 DTEND;TZID=America/Denver:20260925T143000 SUMMARY:Data Science Meets Chemistry - Matthew S. Sigman DESCRIPTION:Matthew S. Sigman (UU Chemistry)\n\nHow do chemists use data sc ience to discover and understand new reactions? This talk offers an access ible introduction to the distinctive challenges of chemical data: small\, complex datasets\; molecular representations\; and the need to connect pre dictions with physical mechanisms. Through examples from reaction developm ent\, catalyst design\, and optimization\, it will show how molecular desc riptors\, statistical modeling\, and machine learning reveal relationships between structure\, reactivity\, and selectivity. Along the way\, the tal k will translate chemical questions for a data science audience\, highligh t opportunities for automation and collaboration\, and consider how new co mputational approaches might expand what chemists can discover.\n\nhttps:/ /datascience.utah.edu/talks/2026-09-25-matthew-s-sigman/ URL:https://datascience.utah.edu/talks/2026-09-25-matthew-s-sigman/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 CATEGORIES:education,optimization END:VEVENT BEGIN:VEVENT UID:2026-09-30-bei-wang-shandian-zhe@datascience.utah.edu DTSTAMP:20260930T180000Z DTSTART;TZID=America/Denver:20260930T120000 DTEND;TZID=America/Denver:20260930T130000 SUMMARY:Structure Without a Grid: Extracting and Tracking Topological Featu res - Bei Wang\, Shandian Zhe DESCRIPTION:Bei Wang (&\, Shandian Zhe)\, Shandian Zhe (&\, Shandian Zhe)\n \nLarge-scale scientific simulations — in fluid dynamics\, climate\,\nco mbustion\, and beyond — are increasingly outpacing our ability to\nstore \, transfer\, and analyze them at full resolution. A growing\nalternative to storing raw discretized grids is the continuous\nimplicit model\, such as multivariate functional approximation (MFA)\nand implicit neural repres entations (INRs). These represent a\nsimulation's fields as continuous fun ctions that can be queried at\narbitrary points and differentiated exactly \, with no resampling\nrequired. This talk presents a line of work buildin g the topological\ndata analysis toolkit needed to make such continuous re presentations\npractically useful. First\, we show how to extract critical points —\nthe maxima\, minima\, and saddles that organize a scalar fiel d's\nstructure — directly from an MFA model\, without ever discretizing back\nto a grid. Second\, we extend this to richer topological descriptors :\ncontours (isosurfaces)\, Jacobi sets (the shared critical structure\nbe tween two co-registered fields\, useful for comparing related\nquantities like pressure and temperature)\, and ridge-valley graphs\,\nwhich trace fi lamentary or ridge-like structure in the data. Third\, we\nintroduce a fra mework for tracking these topological features\ncontinuously through time or parameter space\, following critical\npoints as smooth trajectories usi ng the model's own derivatives rather\nthan resampling frame by frame — avoiding the aliasing artifacts that\nplague discrete feature tracking. To gether\, these methods point toward\na workflow in which large-scale data is stored once as a continuous\nfunctional surrogate\, and structural anal ysis — feature detection\,\ncomparison across fields\, and evolution ove r time — happens directly\non that surrogate\, at whatever resolution th e analysis demands. This\nwork is joint with Guanqun Ma\, David Lenz\, Tom Peterka\, Hanqi Guo\,\nKaiyuan Tang\, and Chaoli Wang.\n\nSpeaker 2: Shan dian Zhe (UU KSoC)\n\nhttps://datascience.utah.edu/talks/2026-09-30-bei-wa ng-shandian-zhe/ URL:https://datascience.utah.edu/talks/2026-09-30-bei-wang-shandian-zhe/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 3780 (Evans) | https://utexas.zoom.us/j/88630218251?pwd=OLbanw 8MQXVnsiho5qaGjdPKxe98Ea.1&jst=1 CATEGORIES:visualization END:VEVENT BEGIN:VEVENT UID:2026-10-02-berton-earnshaw@datascience.utah.edu DTSTAMP:20261002T193000Z DTSTART;TZID=America/Denver:20261002T133000 DTEND;TZID=America/Denver:20261002T143000 SUMMARY:Analyzing -omics Data with High-Dimensional Embeddings: Advantages and Pitfalls - Berton Earnshaw DESCRIPTION:Berton Earnshaw (UU Math+Bio)\n\nHigh-dimensional embeddings of fer a powerful way to summarize complex omics measurements\, such as gene- expression or protein-abundance profiles\, and to compare biological condi tions in a common geometric space. More dimensions can preserve weak or di stributed biological signals that would be missed by a small set of featur es. But additional dimensions also introduce noise\, inflate distances\, a nd can make apparently large differences difficult to interpret.\n\nIn thi s talk\, I will use a simple geometric test based on the squared length of an embedded data vector to illustrate both sides of this tradeoff. We wil l examine how statistical power depends on dimension\, sample size\, mean shifts\, variance changes\, and correlations among features. The analysis reveals several useful principles: weak signals can accumulate across many coordinates\, irrelevant features can reduce sensitivity\, and changes in covariance can alter power even when the average distance remains unchang ed. Along the way\, I will introduce accessible ideas from hypothesis test ing\, Gaussian models\, and high-dimensional geometry. The goal is to prov ide practical intuition for deciding when high-dimensional embeddings help reveal biological perturbations—and when they may instead obscure them. \n\nhttps://datascience.utah.edu/talks/2026-10-02-berton-earnshaw/ URL:https://datascience.utah.edu/talks/2026-10-02-berton-earnshaw/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 CATEGORIES:biology & genomics,deep learning,statistics END:VEVENT BEGIN:VEVENT UID:2026-10-09-david-gleich@datascience.utah.edu DTSTAMP:20261009T180000Z DTSTART;TZID=America/Denver:20261009T120000 DTEND;TZID=America/Denver:20261009T130000 SUMMARY:Powers of magnetic graph matrix: Fourier spectrum\, walk compressio n\, and applications - David Gleich DESCRIPTION:David Gleich (Purdue University)\n\nMagnetic graph matrices are powerful tools for modeling quantum systems and directed networks\, but t heir application in network analysis has been limited by a lack of combina torial understanding. We present a combinatorial interpretation that funda mentally reveals how these matrices encode local network structure. We fur ther show that this structure information is highly compressible in real-w orld networks\, enabling accurate approximations from a small number of ma gnetic potentials. This fresh foundation also unlocks further applications \, like identifying crucial network motifs (e.g.\, frustrated directed cyc les such as feed-forward loops) and enhancing directed link prediction.\n\ nLink to paper:\nhttps://www.pnas.org/doi/10.1073/pnas.2516664123\n\nhttps ://datascience.utah.edu/talks/2026-10-09-david-gleich/ URL:https://datascience.utah.edu/talks/2026-10-09-david-gleich/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2760 | https://utah.zoom.us/j/85983626630 CATEGORIES:algorithms & theory,networks & graphs END:VEVENT BEGIN:VEVENT UID:2026-10-23-kenneth-collins@datascience.utah.edu DTSTAMP:20261023T193000Z DTSTART;TZID=America/Denver:20261023T133000 DTEND;TZID=America/Denver:20261023T143000 SUMMARY:The talk will consider the impact of AI in moving image works\, inc luding AI filmmaking\, animation\, and video art - Kenneth Collins DESCRIPTION:Kenneth Collins (UU Film)\n\nhttps://datascience.utah.edu/talks /2026-10-23-kenneth-collins/ URL:https://datascience.utah.edu/talks/2026-10-23-kenneth-collins/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 CATEGORIES:computer vision END:VEVENT BEGIN:VEVENT UID:2026-10-30-guang-tian@datascience.utah.edu DTSTAMP:20261030T193000Z DTSTART;TZID=America/Denver:20261030T133000 DTEND;TZID=America/Denver:20261030T143000 SUMMARY:Land Use and Transportation Planning with AI - Guang Tian DESCRIPTION:Guang Tian (UU City & Metro Planning + SCI)\n\nhttps://datascie nce.utah.edu/talks/2026-10-30-guang-tian/ URL:https://datascience.utah.edu/talks/2026-10-30-guang-tian/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 CATEGORIES:geospatial END:VEVENT BEGIN:VEVENT UID:2026-11-13-aik-choon-tan@datascience.utah.edu DTSTAMP:20261113T203000Z DTSTART;TZID=America/Denver:20261113T133000 DTEND;TZID=America/Denver:20261113T143000 SUMMARY:AI for Cancer Treatment - Aik Choon Tan DESCRIPTION:Aik Choon Tan (UU Oncological Sciences)\n\nhttps://datascience. utah.edu/talks/2026-11-13-aik-choon-tan/ URL:https://datascience.utah.edu/talks/2026-11-13-aik-choon-tan/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 CATEGORIES:health & medicine END:VEVENT BEGIN:VEVENT UID:2026-12-04-tom-henderson@datascience.utah.edu DTSTAMP:20261204T203000Z DTSTART;TZID=America/Denver:20261204T133000 DTEND;TZID=America/Denver:20261204T143000 SUMMARY:40 Years of AI ! - Tom Henderson DESCRIPTION:Tom Henderson (UU KSoC)\n\nhttps://datascience.utah.edu/talks/2 026-12-04-tom-henderson/ URL:https://datascience.utah.edu/talks/2026-12-04-tom-henderson/ SEQUENCE:0 STATUS:CONFIRMED TRANSP:OPAQUE LOCATION:WEB 2250 | https://utah.zoom.us/j/85983626630 END:VEVENT END:VCALENDAR