What Limits the Performance of Local Self-attention?

lthough self-attention is powerful in modeling long-range dependencies, the performance of local self-attention (LSA) is just similar to depth-wise convolution, which puzzles researchers on whether to use LSA or its counterparts, which one is better, and what limits the performance of LSA. To clarify these, we comprehensively investigate LSA and its counterparts from channel setting and spatial processing . We find that the devil lies in attention generation and application, where relative position embedding and neighboring filter application are key factors. Based on these findings, we propose enhanced local self-attention (ELSA) with Hadamard attention and the ghost head. Hadamard attention introduces the Hadamard product to efficiently generate attention in the neighboring area, while maintaining the high-order mapping. The ghost head combines attention maps with static matrices to increase channel capacity. Experiments demonstrate the effectiveness of ELSA. Without architecture/hyperparameter modification, drop-in replacing LSA with ELSA boosts Swin Transformer by up to + 1.4 on top-1 accuracy. ELSA also consistently benefits VOLO from D1 to D5, where ELSA-VOLO-D5 achieves 87.2 on the ImageNet-1K without extra training images. In addition, we evaluate ELSA in downstream tasks. ELSA significantly improves the baseline by up to + 1.9 box Ap/ + 1.3 mask Ap on the COCO, and by up to + 1.9 mIoU on the ADE20K.