Artifacts and Attention Sinks: Structured Approximations for Efficient Vision Transformers
This work addresses efficiency issues in vision transformers for researchers and practitioners, though it is incremental as it builds on existing approximations.
The paper tackled the problem of inefficient self-attention in vision transformers by analyzing massive and artifact tokens that act as attention sinks, and introduced Fast Nyström Attention (FNA) to approximate self-attention in linear time and space, achieving competitive performance on tasks like retrieval and classification while reducing computational overhead.
Vision transformers have emerged as a powerful tool across a wide range of applications, yet their inner workings remain only partially understood. In this work, we examine the phenomenon of massive tokens - tokens with exceptionally high activation norms that act as attention sinks - and artifact tokens that emerge as a byproduct during inference. Our analysis reveals that these tokens mutually suppress one another through the attention mechanism, playing a critical role in regulating information flow within the network. Leveraging these insights, we introduce Fast Nyström Attention (FNA), a training-free method that approximates self-attention in linear time and space by exploiting the structured patterns formed by massive and artifact tokens. Additionally, we propose a masking strategy to mitigate noise from these tokens, yielding modest performance gains at virtually no cost. We evaluate our approach on popular pretrained vision backbones and demonstrate competitive performance on retrieval, classification, segmentation, and visual question answering (VQA), all while reducing computational overhead.