Systems, methods and techniques for learning and using sparse instance-dependent attention for efficient vision transformers

Abstract

Vision Transformers (ViT) have shown their competitive advantages performance-wise compared to convolutional neural networks (CNNs) though they often come with high computational costs. Methods, systems and techniques herein learn instance-dependent attention patterns, utilizing a lightweight connectivity predictor module to estimate a connectivity score of each pair of tokens. Intuitively, two tokens have high connectivity scores if the features are considered relevant either spatially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature providing an opportunity to accelerate the network via sparse computations. Equipped with the learned unstructured attention pattern, sparse attention ViT produces a superior Pareto-optimal trade-off between FLOPs and top-1 accuracy on ImageNet compared to token sparsity (48%˜69% FLOPs reduction of MHSA; accuracy drop within 0.4%). Combining attention and token sparsity reduces VIT FLOPs by over 60%.