Project Title

Towards Bounding the Behavior of Deep Neural Networks

Academic department under which the project should be listed

CCSE - Computer Science

Faculty Sponsor Name

Arthur Choi

Abstract (300 words maximum)

Recent and rapid advances in Artificial Intelligence (AI), particularly in the form of deep neural networks, have revolutionized a numerous and diverse range of fields. As neural networks become more pervasive, the need to understand the boundaries of their behavior is becoming increasingly important. For example, can we formally guarantee that an autonomous vehicle, controlled by a neural network, will not violate traffic laws, such as reaching excessive speeds? Similarly, can we formally guarantee that an automated gift recommendation system will not propagate and reinforce social biases in the data (e.g., by suggesting doctor costumes for boys and nurse costumes for girls)? In this research, towards the goal of bounding the behavior of a neural network, we propose first to bound the behavior of individual neurons. In particular, we employ a formal type of analysis that allows us to incrementally tighten inner and outer bounds on the behavior of a neuron. In turn, we hope to similarly bound the behavior of a neural network.

Disciplines

Artificial Intelligence and Robotics | Computer Sciences

Project Type

Oral Presentation (15-min time slots)

How will this be presented?

Yes, in person

This document is currently not available here.

Share

COinS
 

Towards Bounding the Behavior of Deep Neural Networks

Recent and rapid advances in Artificial Intelligence (AI), particularly in the form of deep neural networks, have revolutionized a numerous and diverse range of fields. As neural networks become more pervasive, the need to understand the boundaries of their behavior is becoming increasingly important. For example, can we formally guarantee that an autonomous vehicle, controlled by a neural network, will not violate traffic laws, such as reaching excessive speeds? Similarly, can we formally guarantee that an automated gift recommendation system will not propagate and reinforce social biases in the data (e.g., by suggesting doctor costumes for boys and nurse costumes for girls)? In this research, towards the goal of bounding the behavior of a neural network, we propose first to bound the behavior of individual neurons. In particular, we employ a formal type of analysis that allows us to incrementally tighten inner and outer bounds on the behavior of a neuron. In turn, we hope to similarly bound the behavior of a neural network.

blog comments powered by Disqus