Adaptive Gradient Methods with Dynamic Bound of Learning Rate

TL;DR

This paper introduces AdaBound and AMSBound, adaptive optimizers with dynamic learning rate bounds, which smoothly transition from adaptive methods to SGD, improving generalization and training stability.

Contribution

The paper proposes novel variants of Adam and AMSGrad with dynamic bounds on learning rates, providing convergence guarantees and improved performance over existing adaptive methods.

Findings

Eliminates the generalization gap between adaptive methods and SGD

Maintains higher learning speed early in training

Significantly improves performance on complex deep networks

Abstract

Adaptive optimization methods such as AdaGrad, RMSprop and Adam have been proposed to achieve a rapid training process with an element-wise scaling term on learning rates. Though prevailing, they are observed to generalize poorly compared with SGD or even fail to converge due to unstable and extreme learning rates. Recent work has put forward some algorithms such as AMSGrad to tackle this issue but they failed to achieve considerable improvement over existing methods. In our paper, we demonstrate that extreme learning rates can lead to poor performance. We provide new variants of Adam and AMSGrad, called AdaBound and AMSBound respectively, which employ dynamic bounds on learning rates to achieve a gradual and smooth transition from adaptive methods to SGD and give a theoretical proof of convergence. We further conduct experiments on various popular tasks and models, which is often insufficient in previous work. Experimental results show that new variants can eliminate the generalization gap between adaptive methods and SGD and maintain higher learning speed early in training at the same time. Moreover, they can bring significant improvement over their prototypes, especially on complex deep networks. The implementation of the algorithm can be found at https://github.com/Luolc/AdaBound .