Perception as Analysis-by-Synthesis

Tejas Kulkarni is a PhD student at MIT in Josh Tenenbaum's lab and spent last summer working at Google DeepMind in London. His talk will be focused on his recent paper entitled Picture: A Probabilistic Programming Language for Scene Perception (https://mrkulk.github.io/www_cvpr15/1999.pdf)

Abstract:
Recent progress on probabilistic modeling and statistical learning, coupled with the availability of large training datasets, has led to remarkable progress in computer vision. An alternative to the empirical regression based approach, often termed as 'analysis-by-synthesis', instead relies on models or theories of the world, which can be used to interpret perceptual observations. How do we best design systems that can map raw scenes into structured representations? The process of computer graphics provides a flexible conceptual framework to go from structured scene descriptions to images. The problem of vision can be thought of as running this process backwards (therefore termed as 'inverse graphics'). In this talk, Tejas will argue for this perspective and showcase several probabilistic programs and deep neural network architectures on application areas including -- 3D body pose estimation, 3D face analysis, shape from images, and character recognition.