Asymmetric Momentum: A Rethinking of Gradient Descent

TL;DR

This paper introduces Loss-Controlled Asymmetric Momentum (LCAM), a novel SGD enhancement that adaptively accelerates parameters based on loss phases, improving training efficiency and accuracy across datasets without extra computational costs.

Contribution

The paper proposes LCAM, a simple yet effective adaptive momentum method that adjusts to all dataset types by leveraging loss phases, challenging traditional adaptive optimizer assumptions.

Findings

LCAM accelerates slow-changing parameters in sparse gradients.

LCAM accelerates frequently-changing parameters in non-sparse gradients.

LCAM achieves comparable or better accuracy with fewer epochs.

Abstract

Through theoretical and experimental validation, unlike all existing adaptive methods like Adam which penalize frequently-changing parameters and are only applicable to sparse gradients, we propose the simplest SGD enhanced method, Loss-Controlled Asymmetric Momentum(LCAM). By averaging the loss, we divide training process into different loss phases and using different momentum. It not only can accelerates slow-changing parameters for sparse gradients, similar to adaptive optimizers, but also can choose to accelerates frequently-changing parameters for non-sparse gradients, thus being adaptable to all types of datasets. We reinterpret the machine learning training process through the concepts of weight coupling and weight traction, and experimentally validate that weights have directional specificity, which are correlated with the specificity of the dataset. Thus interestingly, we observe that in non-sparse gradients, frequently-changing parameters should actually be accelerated, which is completely opposite to traditional adaptive perspectives. Compared to traditional SGD with momentum, this algorithm separates the weights without additional computational costs. It is noteworthy that this method relies on the network's ability to extract complex features. We primarily use Wide Residual Networks for our research, employing the classic datasets Cifar10 and Cifar100 to test the ability for feature separation and conclude phenomena that are much more important than just accuracy rates. Finally, compared to classic SGD tuning methods, while using WRN on these two datasets and with nearly half the training epochs, we achieve equal or better test accuracy.