cosinegate-dynamic-routing

【效率与协作】Achieve efficient neural networks via self-supervised dynamic routing using Cosine Incompatibility Ratio (CIR). Ground gating decisions in geometric novelty rather than learned heuristics, enable per-sample/per-block binary routing via Gumbel-softmax, constrain with progressive FLOPs regularization—maintaining accuracy while reducing computation 28.5…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准