Mathematics Colloquia and Seminars

Return to Colloquia & Seminar listing

A primary method of implementing constrained stochastic gradient methods relies on projecting back into the feasible region after each gradient step. This projection can be computationally expensive, and is a bottleneck on related algorithms. We explore an alternative way to implement stochastic gradient descent when constraints are present, which is motivated by smoothing techniques detailed in recent work by Beck and Teboulle.