Curved Representation Space of Vision Transformers

Neural networks with self-attention (a.k.a. Transformers) like ViT and Swin have emerged as a better alternative to traditional convolutional neural networks (CNNs). However, our understanding of how the new architecture works is still limited. In this paper, we focus on the phenomenon that Transformers show higher robustness against corruptions than CNNs, while not being overconfident. This is contrary to the intuition that robustness increases with confidence. We resolve this contradiction by empirically investigating how the output of the penultimate layer moves in the representation space as the input data moves linearly within a small area. In particular, we show the following. (1) While CNNs exhibit fairly linear relationship between the input and output movements, Transformers show nonlinear relationship for some data. For those data, the output of Transformers moves in a curved trajectory as the input moves linearly. (2) When a data is located in a curved region, it is hard to move it out of the decision region since the output moves along a curved trajectory instead of a straight line to the decision boundary, resulting in high robustness of Transformers. (3) If a data is slightly modified to jump out of the curved region, the movements afterwards become linear and the output goes to the decision boundary directly. In other words, there does exist a decision boundary near the data, which is hard to find only because of the curved representation space. This explains the underconfident prediction of Transformers. Also, we examine mathematical properties of the attention operation that induce nonlinear response to linear perturbation. Finally, we share our additional findings, regarding what contributes to the curved representation space of Transformers, and how the curvedness evolves during training.