Why Training-Free Token Reduction Collapses: The Inherent Instability of Pairwise Scoring Signals
Uncovering why token reduction methods fail in deep vision models and proposing a more stable alternative.
Training-free token reduction methods for Vision Transformers (ToMe, ToFu, PiToMe, and MCTF) employ different scoring mechanisms, yet they share a closely matched cliff-like collapse at high compression. This paper explains \emph{why}. We develop a diagnostic framework with two t...
ArXiv CS.AI
Source
Generated insight by OnAirToday AI
This research reveals the fundamental limitations of current token reduction methods in Vision Transformers by identifying how unstable pairwise signals lead to catastrophic performance drops. By shifting toward unary signal processing, the study introduces a more robust framework that allows for significant computational efficiency without sacrificing accuracy.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.