Revisiting the Initial Steps in Adaptive Gradient Descent Optimization

TL;DR

This paper identifies the zero initialization of second-order moments in Adam as a key issue and proposes non-zero initialization strategies that improve stability and performance in training deep neural networks.

Contribution

It introduces simple non-zero initialization methods for Adam's second-order moments, significantly improving its stability and generalization in training deep models.

Findings

Non-zero initialization stabilizes Adam training.

Improved final performance with new initialization.

Adam matches recent adaptive optimizer variants.

Abstract

Adaptive gradient optimization methods, such as Adam, are prevalent in training deep neural networks across diverse machine learning tasks due to their ability to achieve faster convergence. However, these methods often suffer from suboptimal generalization compared to stochastic gradient descent (SGD) and exhibit instability, particularly when training Transformer models. In this work, we show the standard initialization of the second-order moment estimation ($v_0 =0$) as a significant factor contributing to these limitations. We introduce simple yet effective solutions: initializing the second-order moment estimation with non-zero values, using either data-driven or random initialization strategies. Empirical evaluations demonstrate that our approach not only stabilizes convergence but also enhances the final performance of adaptive gradient optimizers. Furthermore, by adopting the proposed initialization strategies, Adam achieves performance comparable to many recently proposed variants of adaptive gradient optimization methods. Our code is available at https://github.com/Walleclipse/Adam_Initialization.