Advances in Bounding the Behavior of Neural Networks
Disciplines
Artificial Intelligence and Robotics | Theory and Algorithms
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? Building on our prior results on bounding the behavior of individual neurons, we report progress on bounding the behavior of neural networks. In particular, we show how bounds on multiple neurons can be combined and propagated through a network of logical gates.
Academic department under which the project should be listed
CCSE - Computer Science
Primary Investigator (PI) Name
Arthur Choi
Advances in Bounding the Behavior of 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? Building on our prior results on bounding the behavior of individual neurons, we report progress on bounding the behavior of neural networks. In particular, we show how bounds on multiple neurons can be combined and propagated through a network of logical gates.