Can a person understand how an AI works well enough to reprogram an AI directly?
I will discuss how deep networks can be dissected to reveal the organization of their internal computations. Drawing examples from my work on state-of-the-art deep networks in computer vision, I will show how understanding the neurons of a network can enable a person to modify complex data semantically, for example, by adding or removing objects in a realistic image of a scene. Then I will describe current research to understand and rewrite higher-level rules of generative networks by manipulating the network parameters directly. We will discuss how these examples are first steps toward a new type of human-AI collaborative deep learning, where deep models are cracked open to reveal useful computations about complex data, and where both humans and trained systems contribute and learn new ideas from one another.
David Bau is a PhD student at MIT, advised by Antonio Torralba. His research focuses on the dissection, visualization, and interactive manipulation of deep networks in vision, and he is the creator of the Network Dissection, GAN Paint and GAN Rewriting methods. Previous to MIT he was an engineer at Google where he contributed to Image Search and Hangouts and created the Pencil Code educational programming system. David is coauthor of the widely-used textbook, Numerical Linear Algebra.
Faculty Host: Ameet Talwalkar
Machine Learning Department
Zoom Participation. See announcement.