Scaling Vision Transformers to 22 Billion Parameters
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.
Cite
@article{arxiv.2302.05442,
title = {Scaling Vision Transformers to 22 Billion Parameters},
author = {Mostafa Dehghani and Josip Djolonga and Basil Mustafa and Piotr Padlewski and Jonathan Heek and Justin Gilmer and Andreas Steiner and Mathilde Caron and Robert Geirhos and Ibrahim Alabdulmohsin and Rodolphe Jenatton and Lucas Beyer and Michael Tschannen and Anurag Arnab and Xiao Wang and Carlos Riquelme and Matthias Minderer and Joan Puigcerver and Utku Evci and Manoj Kumar and Sjoerd van Steenkiste and Gamaleldin F. Elsayed and Aravindh Mahendran and Fisher Yu and Avital Oliver and Fantine Huot and Jasmijn Bastings and Mark Patrick Collier and Alexey Gritsenko and Vighnesh Birodkar and Cristina Vasconcelos and Yi Tay and Thomas Mensink and Alexander Kolesnikov and Filip Pavetić and Dustin Tran and Thomas Kipf and Mario Lučić and Xiaohua Zhai and Daniel Keysers and Jeremiah Harmsen and Neil Houlsby},
journal= {arXiv preprint arXiv:2302.05442},
year = {2023}
}