Gated Spatial Redundancy Projection for Pathology Transformer Attentions
British Machine Vision Conference (BMVC 2026), Oral Presentation
GatedSRP is a lightweight, drop-in correction for pathology transformer self-attention. For each patch token and attention head, it estimates a local redundancy direction from neighboring value vectors and learns a signed gate that decides how to correct the aligned component, adding about 0.02% trainable parameters.
