Adaptive momentum with discriminative weight for neural network stochastic optimization

Abstract

Optimization algorithms with momentum have been widely used for building deep learning models because of the fast convergence rate. Momentum helps accelerate Stochastic gradient descent in relevant directions in parameter updating, minifying the oscillations of the parameters update route. The gradient of each step in optimization algorithms with momentum is calculated by a part of the training samples, so there exists stochasticity, which may bring errors to parameter updates. In this case, momentum placing the influence of the last step to the current step with a fixed weight is obviously inaccurate, which propagates the error and hinders the correction of the current step. Besides, such a hyperparameter can be extremely hard to tune in applications as well. In this paper, we introduce a novel optimization algorithm, namely, Discriminative wEight on Adaptive Momentum (DEAM). Instead of assigning the momentum term weight with a fixed hyperparameter, DEAM proposes to compute the momentum weight automatically based on the discriminative angle. The momentum term weight will be assigned with an appropriate value that configures momentum in the current step. In this way, DEAM involves fewer hyperparameters. DEAM also contains a novel backtrack term, which restricts redundant updates when the correction of the last step is needed. The backtrack term can effectively adapt the learning rate and achieve the anticipatory update as well. Extensive experiments demonstrate that DEAM can achieve a faster convergence rate than the existing optimization algorithms in training the deep learning models of both convex and nonconvex situations.