Scaling Laws for Sparsely-Connected Foundation Models

TL;DR

This paper establishes a new scaling law for sparsely-connected foundation models, revealing how sparsity affects performance and optimal configurations across vision and language transformers trained on large datasets.

Contribution

It introduces the first empirical scaling law linking weight sparsity, model size, and training data, with insights into optimal sparsity levels and structures for foundation models.

Findings

Optimal sparsity increases with training data size.

The proposed law is validated across vision and language models.

Different sparsity structures and pretraining strategies are analyzed.

Abstract

We explore the impact of parameter sparsity on the scaling behavior of Transformers trained on massive datasets (i.e., "foundation models"), in both vision and language domains. In this setting, we identify the first scaling law describing the relationship between weight sparsity, number of non-zero parameters, and amount of training data, which we validate empirically across model and data scales; on ViT/JFT-4B and T5/C4. These results allow us to characterize the "optimal sparsity", the sparsity level which yields the best performance for a given effective model size and training budget. For a fixed number of non-zero parameters, we identify that the optimal sparsity increases with the amount of data used for training. We also extend our study to different sparsity structures (such as the hardware-friendly n:m pattern) and strategies (such as starting from a pretrained dense model). Our findings shed light on the power and limitations of weight sparsity across various parameter and computational settings, offering both theoretical understanding and practical implications for leveraging sparsity towards computational efficiency improvements.