学术文献库

Vision transformer (ViT) and its variants have achieved remarkable successes in various visual tasks. The key characteristic of these ViT models is to adopt different aggregation strategies of spatial patch information within the artificial neural networks (ANNs). However, there is still a key lack of unified representation of different ViT architectures for systematic understanding and assessment of model representation performance. Moreover, how those well-performing ViT ANNs are similar to real biological neural networks (BNNs) is largely unexplored. To answer these fundamental questions, we, for the first time, propose a unified and biologically-plausible relational graph representation of ViT models. Specifically, the proposed relational graph representation consists of two key sub-graphs: aggregation graph and affine graph. The former one considers ViT tokens as nodes and describes their spatial interaction, while the latter one regards network channels as nodes and reflects the information communication between channels. Using this unified relational graph representation, we found that: a) a sweet spot of the aggregation graph leads to ViTs with significantly improved predictive performance; b) the graph measures of clustering coefficient and average path length are two effective indicators of model prediction performance, especially when applying on the datasets with small samples; c) our findings are consistent across various ViT architectures and multiple datasets; d) the proposed relational graph representation of ViT has high similarity with real BNNs derived from brain science data. Overall, our work provides a novel unified and biologically-plausible paradigm for more interpretable and effective representation of ViT ANNs.