You Don't Need Domain-Specific Data Augmentations When Scaling Self-Supervised Learning

TL;DR

This paper demonstrates that large-scale self-supervised learning with joint-embedding architectures can achieve state-of-the-art results using minimal data augmentations, challenging the necessity of traditional augmentation techniques.

Contribution

The study shows that at scale, strong image representations can be learned with minimal augmentations, specifically only cropping, which questions the established belief in the importance of data augmentation in SSL.

Findings

Strong representations achieved with only cropping augmentation.

State-of-the-art results obtained with minimal augmentations.

Compute constraints influence experimental conclusions.

Abstract

Self-Supervised learning (SSL) with Joint-Embedding Architectures (JEA) has led to outstanding performances. All instantiations of this paradigm were trained using strong and well-established hand-crafted data augmentations, leading to the general belief that they are required for the proper training and performance of such models. On the other hand, generative reconstruction-based models such as BEIT and MAE or Joint-Embedding Predictive Architectures such as I-JEPA have shown strong performance without using data augmentations except masking. In this work, we challenge the importance of invariance and data-augmentation in JEAs at scale. By running a case-study on a recent SSL foundation model - DINOv2 - we show that strong image representations can be obtained with JEAs and only cropping without resizing provided the training data is large enough, reaching state-of-the-art results and using the least amount of augmentation in the literature. Through this study, we also discuss the impact of compute constraints on the outcomes of experimental deep learning research, showing that they can lead to very different conclusions.