What are you sinking? A geometric approach on attention sink
This provides foundational insights into transformer attention mechanisms, potentially influencing architecture design for researchers and practitioners in machine learning.
The paper demonstrates that attention sink patterns in transformers are not architectural artifacts but emerge from fundamental geometric principles of establishing reference frames in high-dimensional spaces, identifying three distinct reference frame types that correlate with attention sink phenomena.
Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens (often special tokens or positional anchors) disproportionately attract attention from other tokens. We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle: the establishment of reference frames that anchor representational spaces. We analyze several architectures and identify three distinct reference frame types, centralized, distributed, and bidirectional, that correlate with the attention sink phenomenon. We show that they emerge during the earliest stages of training as optimal solutions to the problem of establishing stable coordinate systems in high-dimensional spaces. We show the influence of architecture components, particularly position encoding implementations, on the specific type of reference frame. This perspective transforms our understanding of transformer attention mechanisms and provides insights for both architecture design and the relationship with AS.