Stochastic Weight Averaging Revisited

Abstract

Averaging neural network weights sampled by a backbone stochastic gradient descent (SGD) is a simple-yet-effective approach to assist the backbone SGD in finding better optima, in terms of generalization. From a statistical perspective, weight-averaging contributes to variance reduction. Recently, a well-established stochastic weight-averaging (SWA) method was proposed, which featured the application of a cyclical or high-constant (CHC) learning-rate schedule for generating weight samples for weight-averaging. Then, a new insight on weight-averaging was introduced, which stated that weight average assisted in discovering a wider optima and resulted in better generalization. We conducted extensive experimental studies concerning SWA, involving 12 modern deep neural network model architectures and 12 open-source image, graph, and text datasets as benchmarks. We disentangled the contributions of the weight-averaging operation and the CHC learning-rate schedule for SWA, showing that the weight-averaging operation in SWA still contributed to variance reduction, and the CHC learning-rate schedule assisted in exploring the parameter space more widely than the backbone SGD, which could be be under-fitted due to a lack of training budget. We then presented an algorithm termed periodic SWA (PSWA) that comprised a series of weight-averaging operations to exploit such wide parameter space structures as explored by the CHC learning-rate schedule, and we empirically demonstrated that PSWA outperformed its backbone SGD remarkably.

Keywords

• deep neural network
• stochastic gradient descent
• stochastic weight-averaging
• generalization
• wide optima
• learning rate