المفاهيم الأساسية
SGD-exp provides nearly linear convergence guarantees for Massart noise corruption in streaming linear and ReLU regression.
الإحصائيات
"We show novel nearly linear convergence guarantees of SGD-exp to the true parameter with up to 50% Massart corruption rate."
"Our analysis also reveals that SGD-exp tolerates any corruption probability less than 1 when the corruption is symmetric oblivious noise."
"The exponential step decay scheduling for SGD are commonly used as a default setting in many popular machine learning software packages."
اقتباسات
"Our analysis is based on the drift analysis of a discrete stochastic process, which could also be interesting on its own."
"Essentially, it implies that in the worst case a presumed adversary can examine all past equations to choose the worst possible placement of corrupted measurements."