Self-attention in vision transformers performs perceptual grouping, not attention

Abstract

Recently, a considerable number of studies in computer vision involve deep neural architectures called vision transformers. Visual processing in these models incorporates computational models that are claimed to implement attention mechanisms. Despite an increasing body of work that attempts to understand the role of attention mechanisms in vision transformers, their effect is largely unknown. Here, we asked if the attention mechanisms in vision transformers exhibit similar effects as those known in human visual attention. To answer this question, we revisited the attention formulation in these models and found that despite the name, computationally, these models perform a special class of relaxation labeling with similarity grouping effects. Additionally, whereas modern experimental findings reveal that human visual attention involves both feed-forward and feedback mechanisms, the purely feed-forward architecture of vision transformers suggests that attention in these models cannot have the same effects as those known in humans. To quantify these observations, we evaluated grouping performance in a family of vision transformers. Our results suggest that self-attention modules group figures in the stimuli based on similarity of visual features such as color. Also, in a singleton detection experiment as an instance of salient object detection, we studied if these models exhibit similar effects as those of feed-forward visual salience mechanisms thought to be utilized in human visual attention. We found that generally, the transformer-based attention modules assign more salience either to distractors or the ground, the opposite of both human and computational salience. Together, our study suggests that the mechanisms in vision transformers perform perceptual organization based on feature similarity and not attention.