MaiT: integrating spatial locality into image transformers with attention masks

Though image transformers have shown competitive results with convolutional neural networks in computer vision tasks, lacking inductive biases such as locality still poses problems in terms of model efficiency especially for embedded applications. In this work, we address this issue by introducing attention masks to incorporate spatial locality into self-attention heads of transformers. Local dependencies are captured with masked attention heads along with global dependencies captured by original unmasked attention heads. With Masked attention image Transformer – MaiT, top-1 accuracy increases by up to 1.0\% compared to DeiT, without extra parameters, computation, or external training data. Moreover, attention masks regulate the training of attention maps, which facilitates the convergence and improves the accuracy of deeper transformers. Masked attention heads guide the model to focus on local information in early layers and promote diverse attention maps in latter layers. Deep MaiT improves the top-1 accuracy by up to 1.5\% compared to CaiT with fewer parameters and less FLOPs. Encoding locality with attention masks requires no extra parameter or structural change, and thus it can be combined with other techniques for further improvement in vision transformers.