aSGD: Stochastic Gradient Descent with Adaptive Batch Size for Every Parameter

Abstract

In recent years, deep neural networks (DNN) have been widely used in many fields. Lots of effort has been put into training due to their numerous parameters in a deep network. Some complex optimizers with many hyperparameters have been utilized to accelerate the process of network training and improve its generalization ability. It often is a trial-and-error process to tune these hyperparameters in a complex optimizer. In this paper, we analyze the different roles of training samples on a parameter update, visually, and find that a training sample contributes differently to the parameter update. Furthermore, we present a variant of the batch stochastic gradient decedent for a neural network using the ReLU as the activation function in the hidden layers, which is called adaptive stochastic gradient descent (aSGD). Different from the existing methods, it calculates the adaptive batch size for each parameter in the model and uses the mean effective gradient as the actual gradient for parameter updates. Experimental results over MNIST show that aSGD can speed up the optimization process of DNN and achieve higher accuracy without extra hyperparameters. Experimental results over synthetic datasets show that it can find redundant nodes effectively, which is helpful for model compression.