Stochastic gradient descent for streaming linear and rectified linear systems with adversarial corruptions
This addresses robust learning in streaming data for applications like real-time analytics, though it is incremental as it builds on existing SGD techniques with a new step size.
The paper tackles robust linear and ReLU regression under adversarial corruptions in a streaming setting, achieving nearly linear convergence to the true parameter with up to 50% corruption rates and improved rates over prior methods.
We propose SGD-exp, a stochastic gradient descent approach for linear and ReLU regressions under Massart noise (adversarial semi-random corruption model) for the fully streaming setting. We show novel nearly linear convergence guarantees of SGD-exp to the true parameter with up to $50\%$ Massart corruption rate, and with any corruption rate in the case of symmetric oblivious corruptions. This is the first convergence guarantee result for robust ReLU regression in the streaming setting, and it shows the improved convergence rate over previous robust methods for $L_1$ linear regression due to a choice of an exponentially decaying step size, known for its efficiency in practice. Our analysis is based on the drift analysis of a discrete stochastic process, which could also be interesting on its own.