Structural Abstractions of Neural Networks – A Talk by Diganta Mukhopadhyay

About the Talk
Neural networks are being used extensively in several safety critical domains, which makes the formal verification of neural networks being deployed in safety critical domains extremely important. However, the size of neural networks remains a major hurdle for verification of neural networks. Structural abstraction of neural networks attempts to solve this problem by starting with a query involving a large network and reducing it to an over-approximate query involving a smaller abstract network. In this talk, Diganta Mukhopadhyay will introduce a wide range of works performing neural network abstraction and present the techniques being used there. Finally, the speaker will present a unifying framework we have developed that brings together several existing approaches for neural network abstraction.

About the Speaker
Diganta Mukhopadhyay is a Researcher at TCS Research, deeply interested in the intersection of formal methods and machine learning. He completed his BSc in Mathematics and Computer Science, and his MSc in Computer Science from Chennai Mathematical Institute. He is currently exploring methods to improve the scalability of verification of deep neural networks.

This event is open to all.

Click to Attend