Scaling Vision Transformers to 22 Billion Parameters

TL;DR

This paper introduces a highly efficient method for training a 22-billion-parameter Vision Transformer, demonstrating improved performance, fairness, robustness, and alignment with human perception, marking a significant step towards scaling vision models.

Contribution

We present a novel recipe for training extremely large Vision Transformers, achieving 22B parameters and demonstrating benefits similar to large language models in vision tasks.

Findings

ViT-22B shows improved performance with scale

Enhanced fairness and robustness observed

Achieves state-of-the-art shape/texture bias alignment

Abstract

The scaling of Transformers has driven breakthrough capabilities for language models. At present, the largest large language models (LLMs) contain upwards of 100B parameters. Vision Transformers (ViT) have introduced the same architecture to image and video modelling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters (Chen et al., 2022). We present a recipe for highly efficient and stable training of a 22B-parameter ViT (ViT-22B) and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features), ViT-22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT-22B demonstrates the potential for "LLM-like" scaling in vision, and provides key steps towards getting there.