Data Science & AI Lecture Series
Provable Editing of Deep Neural Networks
Aditya Thakur
Upcoming
Provable Editing of Deep Neural Networks
| When | Wednesday, October 21, 2026, 3:30 PM – 4:30 PM (MT) |
|---|---|
| Where | 3147 MEB |
Abstract
There are many applications where deep neural networks (DNNs) need to satisfy constraints, e.g., safety-critical domains or scientific machine learning. Testing and verification can determine whether DNN satisfies a constraint. This talk addresses the next step of this process: what to do once faulty behavior is identified. I will discuss recent advances in the problem of provable editing of DNNs. Given a trained DNN and a specification, provable editing makes small changes to the parameters of the DNN to guarantee that the edited DNN satisfies the given specification while still ensuring high accuracy. I will discuss efficient algorithms for this problem that support a wide range of DNN architectures and edit specifications. Our techniques have been applied to ACAS Xu, ImageNet ResNet DNNs, Vision Transformers, Llama, and Qwen models.
Speaker
Aditya Thakur
University of California, Davis
Aditya Thakur is an Associate Professor in the Department of Computer Science at the University of California, Davis. He received his Ph.D. from the University of Wisconsin–Madison. His research interests include programming languages, machine learning, formal methods, and software engineering. He was the recipient of a NSF CAREER award, DOE Early Career Award, Facebook Probability and Programming Research Awards (2019 and 2020), and a Facebook Testing and Verification Research Award.
Tags: deep learning machine learning
Part of the Data Science & AI Lecture Series. Something wrong on this page? Edit _data/talks/2026-10-21-aditya-thakur.toml.