Events

CS Colloquium: Trustworthy Deep Learning: methods, systems and theory

Speaker: Matthew Mirman

Location: 60 Fifth Avenue, Room 150
Videoconference link: https://nyu.zoom.us/j/95135973080

Date: Wednesday, March 23, 2022

Deep learning models are quickly becoming an integral part of a plethora
of high stakes applications, including autonomous driving and health
care. As the discovery of vulnerabilities and flaws in these models has
become frequent, so has the interest in ensuring their safety,
robustness and reliability. My research addresses this need by
introducing new core methods and systems that can establish desirable
mathematical guarantees of deep learning models.

In the first part of my talk I will describe how we leverage abstract
interpretation to scale verification to orders of magnitude larger deep
neural networks than prior work, at the same time demonstrating the
correctness of significantly more properties. I will then show how
these techniques can be extended to ensure, for the first time, formal
guarantees of probabilistic semantic specifications using generative models.

In the second part, I will show how to fuse abstract interpretation with
the training phase so as to improve a model’s amenability to
certification, allowing us to guarantee orders of magnitude more
properties than possible with prior work. Finally, I will discuss
exciting theoretical advances which address fundamental questions on the
very existence of certified deep learning.