CAM Colloquium: Bharath Hariharan (CS, Cornell) - The enduring mysteries of learned visual representations

Description

Abstract:
Object recognition in computer vision has seen a surge of progress in the past half-decade, powered by deep “convolutional neural networks.” One of the most under-appreciated yet surprising findings from this progress is that convolutional networks trained on one task learn internal representations of images that are useful for a completely different task. This kind of “meta”-generalization goes above and beyond what is traditionally explored in machine learning. As such, it opens up several intriguing questions: why does such generalization happen, what properties of the “source” and “target” task affect it, and how can we improve this generalization? In this talk, I will present results from my group as well as the broader computer vision community that has begun to explore these questions.

Bio:
Bharath Hariharan is an assistant professor in the Department of Computer Science at Cornell University. He came to Cornell after a 2-year stint at Facebook AI Research, before which he did his PhD at UC Berkeley. His interests are in computer vision, specifically on visual recognition: designing machine vision systems that can recognize objects as quickly and effectively as humans do.