Student Seminars

Chansoo Lee
PhD student, Department of Electrical Engineering, University of Michigan

Title: Duality between regularizers and stochastic perturbations

Regularization via perturbation techniques have become popular among machine learning practitioners because generic perturbations (e.g. dropout or Gaussian noise) often work just as well as carefully chosen convex regularizers. We show that in online linear optimization, adding Gaussian noise to data is the dual of adding a strongly convex regularization penalty to the loss function. Based on observation, we prove that adding Gaussian noise to data produces low-regret algorithms for two canonical online linear optimization settings.

Student Seminar Archive

For questions regarding the Statistics Student Seminar or if you are interested in presenting, please contact Joonha Park(joonhap@umich.edu) or Jingshen Wang(jshwang@umich.edu).