CS Colloquium
Speaker: Michael Maire, TTIC
Title: Architecting Deep Convolutional Neural Networks for Computer Vision
Host: Holly Rushmeier
Abstract:
Deep convolutional neural networks (CNNs) are central to modern computer vision systems. In this talk, I will present recent work exploring ideas about both CNN architectures and training procedures. I will connect novel architectural design principles with specific capabilities exhibited by networks implementing them. Of particular importance is a multigrid extension of CNNs, in which network layers operate across scale space. Multigrid structure enables networks to learn internal attention and dynamic routing mechanisms, and use them to accomplish visual tasks which standard CNNs fail. On the training side, I will discuss pathways for scaling visual learning beyond current supervised approaches. Self-supervision, by deriving informative tasks from unlabeled data, provides one such route. In addition to our work in this area, I will also discuss a new regularization technique that squeezes more information from detailed and structured labels when training CNNs in a supervised fashion.
Bio:
Michael Maire is a Research Assistant Professor at the Toyota Technological Institute at Chicago (TTIC). He received a B.S. with honors from the California Institute of Technology (Caltech) in 2003, and a Ph.D. in computer science from the University of California, Berkeley, in 2009. Prior to joining TTIC, he was a postdoctoral scholar in the Department of Electrical Engineering at Caltech.