Heavy-Tailed Class Imbalance and Why Adam Outperforms Gradient Descent on Language Models
This addresses a key performance gap in training language models, which is incremental as it explains an existing phenomenon rather than introducing a new method.
The paper tackles the problem of why Adam outperforms gradient descent on large language models more than on other tasks, showing that heavy-tailed class imbalance in language data causes gradient descent to converge slowly on infrequent words, while Adam is less sensitive, with empirical evidence across architectures and a theoretical proof in continuous time.
Adam has been shown to outperform gradient descent on large language models by a larger margin than on other tasks, but it is unclear why. We show that a key factor in this performance gap is the heavy-tailed class imbalance found in language tasks. When trained with gradient descent, the loss of infrequent words decreases more slowly than the loss of frequent ones. This leads to a slow decrease on the average loss as most samples come from infrequent words. On the other hand, Adam and sign-based methods are less sensitive to this problem. To establish that this behavior is caused by class imbalance, we show empirically that it can be reproduced across architectures and data types, on language transformers, vision CNNs, and linear models. On a linear model with cross-entropy loss, we show that class imbalance leads to imbalanced, correlated gradients and Hessians that have been hypothesized to benefit Adam. We also prove that, in continuous time, gradient descent converges slowly on low-frequency classes while sign descent does not.