Generalization Properties and Implicit Regularization for Multiple Passes SGM
This work addresses generalization in machine learning for researchers, but it is incremental as it builds on existing stability and regularization concepts.
The paper tackles the generalization properties of stochastic gradient methods for convex loss functions and linearly parameterized models, showing that step-size or number of passes can control stability and approximation as implicit regularization, with numerical results supporting the theory.
We study the generalization properties of stochastic gradient methods for learning with convex loss functions and linearly parameterized functions. We show that, in the absence of penalizations or constraints, the stability and approximation properties of the algorithm can be controlled by tuning either the step-size or the number of passes over the data. In this view, these parameters can be seen to control a form of implicit regularization. Numerical results complement the theoretical findings.