Oral
Characterizing Implicit Bias in Terms of Optimization Geometry
Suriya Gunasekar · Jason Lee · Daniel Soudry · Nati Srebro

Thu Jul 12th 01:50 -- 02:00 PM @ A9

We study the bias of generic optimization methods, including MirrorDescent, Natural Gradient Descent and Steepest Descent with respectto different potentials and norms, when optimizing underdeterminedlinear models or separable linear classification problems. We askthe question of whether the global minimum (among the many possibleglobal minima) reached by optimization can be characterized in termsof the potential or norm, and indecently of hyper-parameter choices,such as stepsize and momentum.

Author Information

Suriya Gunasekar (Toyota Technological Institute at Chicago)
Jason Lee (University of Southern California)
Daniel Soudry (Technion)
Nati Srebro (Toyota Technological Institute at Chicago)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors