Reducing Bias in Deep Learning Optimization: The RSGDM Approach

Abstract

Currently, widely used first-order deep learning optimizers include non-adaptive learning rate optimizers and adaptive learning rate optimizers. The former is represented by SGDM (Stochastic Gradient Descent with Momentum), while the latter is represented by Adam. Both of these methods use exponential moving averages to estimate the overall gradient. However, estimating the overall gradient using exponential moving averages is biased and has a lag. This paper proposes an RSGDM algorithm based on differential correction. Our contributions are mainly threefold: 1) Analyze the bias and lag brought by the exponential moving average in the SGDM algorithm. 2) Use the differential estimation term to correct the bias and lag in the SGDM algorithm, proposing the RSGDM algorithm. 3) Experiments on the CIFAR datasets have proven that our RSGDM algorithm is superior to the SGDM algorithm in terms of convergence accuracy.