Towards Bounding the Behavior of Deep Neural Networks
Abstract (300 words maximum)
Advances in Artificial Intelligence (AI), particularly in the form of deep neural networks, have revolutionized a 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 will not violate traffic laws, such as reaching excessive speeds? Towards the goal of bounding the behavior of a neural network, we propose how to bound the behavior of individual neurons by incrementally tightening formal bounds on it. We further provide a case study on classifying handwritten digits to illustrate the utility of our algorithm in terms of bounding the behavior of an individual neuron.
Academic department under which the project should be listed
Computer Science
Primary Investigator (PI) Name
Arthur Choi
Towards Bounding the Behavior of Deep Neural Networks
Advances in Artificial Intelligence (AI), particularly in the form of deep neural networks, have revolutionized a 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 will not violate traffic laws, such as reaching excessive speeds? Towards the goal of bounding the behavior of a neural network, we propose how to bound the behavior of individual neurons by incrementally tightening formal bounds on it. We further provide a case study on classifying handwritten digits to illustrate the utility of our algorithm in terms of bounding the behavior of an individual neuron.