Do Vision Transformers See Like Convolutional Neural Networks?

TL;DR

This paper compares Vision Transformers and CNNs, revealing fundamental differences in their internal representations, the role of self-attention and residuals, and implications for spatial localization and transfer learning.

Contribution

It provides a detailed analysis of how ViTs differ from CNNs internally, highlighting the impact of self-attention and residual connections on their representations and spatial information preservation.

Findings

ViTs have more uniform representations across layers.

Self-attention enables early global information aggregation.

ViTs preserve input spatial information effectively.

Abstract

Convolutional neural networks (CNNs) have so far been the de-facto model for visual data. Recent work has shown that (Vision) Transformer models (ViT) can achieve comparable or even superior performance on image classification tasks. This raises a central question: how are Vision Transformers solving these tasks? Are they acting like convolutional networks, or learning entirely different visual representations? Analyzing the internal representation structure of ViTs and CNNs on image classification benchmarks, we find striking differences between the two architectures, such as ViT having more uniform representations across all layers. We explore how these differences arise, finding crucial roles played by self-attention, which enables early aggregation of global information, and ViT residual connections, which strongly propagate features from lower to higher layers. We study the ramifications for spatial localization, demonstrating ViTs successfully preserve input spatial information, with noticeable effects from different classification methods. Finally, we study the effect of (pretraining) dataset scale on intermediate features and transfer learning, and conclude with a discussion on connections to new architectures such as the MLP-Mixer.