Joint merging and pruning: adaptive selection of better token compression strategy

Vision transformer (ViT) is widely used to handle artificial intelligence tasks, making significant advances in a variety of computer vision tasks. However, due to the secondary interaction between tokens, the ViT model is inefficient, which greatly limits the application of the ViT model in real scenarios. In recent years, people have noticed that not all tokens contribute equally to the final prediction of the model, so token compression methods have been proposed, which are mainly divided into token pruning and token merging. Yet, we believe that neither pruning only to reduce non-critical tokens nor merging to reduce similar tokens are optimal strategies for token compression. To overcome this challenge, this work proposes a token compression framework: joint merging and pruning (JMP), which adaptively selects a better token compression strategy based on the similarity between critical tokens and non-critical tokens in each sample. JMP effectively reduces computational complexity while maintaining model performance and does not require the introduction of additional trainable parameters, achieving a good balance between efficiency and performance. Taking DeiT-S as an example, JMP reduces floating point operations by 35% and increases throughput by more than 45% while only decreasing accuracy by 0.2% on ImageNet.