Sparse Double Descent in Vision Transformers: real or phantom threat?

TL;DR

This paper investigates whether Vision Transformers experience the sparse double descent phenomenon and finds that proper regularization can mitigate it, though with trade-offs in model compression.

Contribution

The study demonstrates that optimal regularization can prevent sparse double descent in ViTs, highlighting a practical approach to improve their generalization.

Findings

Optimal $\\ell_2$ regularization relieves sparse double descent in ViTs.

Proper tuning of regularization sacrifices some model compression.

ViTs are not inherently prone to sparse double descent with correct regularization.

Abstract

Vision transformers (ViT) have been of broad interest in recent theoretical and empirical works. They are state-of-the-art thanks to their attention-based approach, which boosts the identification of key features and patterns within images thanks to the capability of avoiding inductive bias, resulting in highly accurate image analysis. Meanwhile, neoteric studies have reported a ``sparse double descent'' phenomenon that can occur in modern deep-learning models, where extremely over-parametrized models can generalize well. This raises practical questions about the optimal size of the model and the quest over finding the best trade-off between sparsity and performance is launched: are Vision Transformers also prone to sparse double descent? Can we find a way to avoid such a phenomenon? Our work tackles the occurrence of sparse double descent on ViTs. Despite some works that have shown that traditional architectures, like Resnet, are condemned to the sparse double descent phenomenon, for ViTs we observe that an optimally-tuned $\ell_2$ regularization relieves such a phenomenon. However, everything comes at a cost: optimal lambda will sacrifice the potential compression of the ViT.