Blurring the Line between Structure and Learning for Adaptive Local Recognition - Evan Shelhamer (UC Berkeley)

Date
Location

ICCSX836

Host: Mark Schmidt

Title:Blurring the Line between Structure and Learning for Ada ptive Local Recognition

Abstract: The visual world is vast and varied , but there is nevertheless ubiquitous structure. In this talk I will foc us on incorporating locality and scale structure into deep networks for im age-to-image tasks, and then examine how dynamic inference that adapts th ese structures to each input helps cope with variability. I will look at t hese directions through the lens of local recognition tasks that require i nference of what and where. By composing structured Gaussian filters with free-form filters, and learning both, our approach optimizes and adapts receptive field size. In effect this controls the degree of locality durin g learning and inference: changes in our parameters would require changes in architecture for standard networks. Multi-step adaptivity, through gra dient optimization of scale during inference, further improves accuracy a nd robustness. This kind of factorization points to a reconciliation of st ructure and learning, through which known visual structure is respected a nd unknown visual detail is learned freely.

Bio: Evan Shelhamer is a hot-off-the-press PhD from UC Berkeley advised by Trevor Darrell. His rese arch focuses on computer vision and machine learning, in particular makin g visual structure differentiable and inference adaptive. His joint work o n fully convolutional networks won best paper honorable mention at CVPR'15 . He was the lead developer of the Caffe deep learning framework from vers ion 0.1 to 1.0, and shared the Mark Everingham service award for Caffe atICCV'17. Before Berkeley, he studied computer science (AI concentration)and psychology at University of Massachusetts Amherst advised by Erik Lea rned-Miller. He takes his coffee black.

If you wish to meet with Evan Shelhamer, please contact Kath Imhiran at lci-admin@cs.ubc.ca.