Yuki Igaue

Hiroaki Aizawa, Yuki Igaue

Transformers are strong baselines in both vision and language because self-attention captures long-range dependencies across tokens. However, the cost of self-attention grows quadratically with the number of tokens. Patch pruning mitigates this cost by estimating per-patch importance and removing redundant patches. To identify informative patches for pruning, we introduce a criterion based on the Shannon entropy of the attention distribution. Low-entropy patches, which receive selective and concentrated attention, are kept as important, while high-entropy patches with attention spread across many locations are treated as redundant. We also extend the criterion from Shannon to Rényi entropy, which emphasizes sharp attention peaks and supports pruning strategies that adapt to task needs and computational limits. In experiments on fine-grained image recognition, where patch selection is critical, our method reduced computation while preserving accuracy. Moreover, adjusting the pruning policy through the Rényi entropy measure yields further gains and improves the trade-off between accuracy and computation.

Yuki Igaue, Hiroaki Aizawa

Multi-head self-attention is a distinctive feature extraction mechanism of vision transformers that computes pairwise relationships among all input patches, contributing significantly to their high performance. However, it is known to incur a quadratic computational complexity with respect to the number of patches. One promising approach to address this issue is patch pruning, which improves computational efficiency by identifying and removing redundant patches. In this work, we propose a patch pruning strategy that evaluates the importance of each patch based on the variance of attention weights across multiple attention heads. This approach is inspired by the design of multi-head self-attention, which aims to capture diverse attention patterns across different subspaces of feature representations. The proposed method can be easily applied during both training and inference, and achieves improved throughput while maintaining classification accuracy in scenarios such as fine-tuning with pre-trained models. In addition, we also found that using robust statistical measures, such as the median absolute deviation in place of variance, to assess patch importance can similarly lead to strong performance. Furthermore, by introducing overlapping patch embeddings, our method achieves better performance with comparable throughput to conventional approaches that utilize all patches.