Learning Geometry, and the Geometry of Learning

Stefano Soatto
University of California, Los Angeles (UCLA)

I will discuss two areas where geometry enters deep learning: One is where we try to use information from large datasets to learn priors for the connectivity/topology of point clouds. The other where we analyze the dynamics of stochastic gradient descent (SGD) to determine the high-dimensional geometry of the loss landscape of deep networks, that reveals “bottlenecks”, “corridors”, and limit cycles.

Back to Workshop II: Shape Analysis