Domain-independent Dominance of Adaptive Methods

TL;DR

This paper introduces AvaGrad, a new adaptive optimizer that outperforms SGD on vision tasks by decoupling learning rate and adaptability, and shows Adam can also excel when properly tuned, challenging conventional beliefs.

Contribution

The paper presents AvaGrad, a novel optimizer with improved performance on vision tasks, and reveals the importance of decoupling learning rate from adaptability for better hyperparameter tuning.

Findings

AvaGrad outperforms SGD on vision tasks with proper tuning.

Decoupling learning rate and adaptability simplifies hyperparameter search.

Adam can outperform SGD on vision tasks when coupling is properly managed.

Abstract

From a simplified analysis of adaptive methods, we derive AvaGrad, a new optimizer which outperforms SGD on vision tasks when its adaptability is properly tuned. We observe that the power of our method is partially explained by a decoupling of learning rate and adaptability, greatly simplifying hyperparameter search. In light of this observation, we demonstrate that, against conventional wisdom, Adam can also outperform SGD on vision tasks, as long as the coupling between its learning rate and adaptability is taken into account. In practice, AvaGrad matches the best results, as measured by generalization accuracy, delivered by any existing optimizer (SGD or adaptive) across image classification (CIFAR, ImageNet) and character-level language modelling (Penn Treebank) tasks.