Building Vision Systems with Modern Machine Learning

Vision appears easy for humans but has proven exceedingly difficult to solve with computers. This seeming paradox has highlighted the surprisingly rich mathematics required to understand the content of visual imagery. Recent advances in machine learning have altered this equation substantially and transformed our expectations of how computers understand the visual world. We are now able to build models that capture complex statistical relationships of visual imagery and exploit these models to automatically describe content — if not synthesize visual content de novo. In this talk I will describe selected research projects of myself and colleagues that have built state-of-the-art models to address concrete problems in vision. In particular, I will focus on how our understanding of vision has been enriched by importing ideas from disparate fields including natural language processing and game theory. In each of these vignettes, my hope is to convey the excitement and potential for enriching our understanding of vision by leveraging modern advances in machine learning.
Jon Shlens is a senior research scientist at Google. Prior to joining Google Research in 2010, he received his Ph.D. in computational neuroscience at UC San Diego and the Salk Institute. He continued his training as a research fellow at the Howard Hughes Medical Institute and a Miller Fellow at UC Berkeley. His research interests include machine perception, statistical signal processing, machine learning and computational neuroscience.