Jul 2026
Brain-Aligned Multi-Stream Video Transformers with Sparse Self-Selection
A sparse winner-takes-all token selection module is introduced that replaces dense self-attention to improve efficiency and approximate competitive routing observed in biological visual circuits and proposes a neuro-inspired split-and-fuse video transformer.
AmirHosein Fadaei, Mahyar Maleki, M. Dehaqani
· arXiv.org · 0 citations