A0824
Title: A robust stochastic method for linear/ReLU regressions under adversarial noise in the streaming setting
Authors: Halyun Jeong - The State University of New York at Albany (United States) [presenting]
Abstract: SGD-exp, a novel stochastic gradient descent method, is introduced with an exponentially decaying step size for linear and ReLU regressions in adversarial corruption models. SGD-exp is designed specifically for fully streaming data environments, where past data cannot be revisited, posing significant challenges for conventional robust regression methods. Near-linear convergence guarantees to the true parameter are presented, even with corruption rates as high as for Massart noise, a semi-random noise model, and it remains effective under any corruption rate for symmetric oblivious corruptions. The analysis is based on the drift analysis of a discrete stochastic process and its hitting time, combined with an SGD residual equation transformation, which could be interesting on its own.
