Seminar by Matus Telgarsky

Probability seminar series

Matus Telgarsky
New York University

Room: M3 3127

Adam and friends via duality

This talk shows how various optimization methods can be derived and analyzed via a duality framework.  Superficially this seems nice due to faster convergence rates than non-duality proofs, but the real value is that unifying in the dual highlights a different set of similarities than in the primal.  The work will primarily be in unpublishable settings (linear separability), but will use slightly less unpublishiable settings (deep networks from ten years ago) for motivation.

The talk will also feature a brief general-audience discussion of some issues facing the ml theory and general academic communities.

Joint work with Ziwei Ji, Danny Son, Zihan Wang.