Profile My skills Saved Collections Edit profile Submit a skill

Cosinegate Dynamic Routing

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% on CIFAR-10.

adu2021 08a69a1 1.1 KB Updated
Coding & Dev Tools

File contents

Overview

Dynamic routing based on semantic novelty between identity and residual paths.

Core Technique

Cosine Incompatibility Ratio:

# CIR = 1 - cos(x, F(x))
# Low CIR: identity sufficient (skip block)
# High CIR: block adds novelty (execute)

cir_score = 1.0 - cosine_similarity(input_x, residual_output)
should_skip = cir_score < threshold

Differentiable Gating: Gumbel-softmax for training, deterministic for inference.

Performance

  • 91.3% accuracy with 28.5% FLOPs saving
  • No auxiliary supervision needed

References

  • Cosine incompatibility ratio for routing
  • Geometric grounding of gate decisions
  • FLOPs-constrained gating

adu2021/skillxiv/tree/main/skills/skillxiv-v0.0.2-claude-opus-4.6/cosinegate-dynamic-routing commit 08a69a1aee

Browse category

Frequently asked questions

npx skillmds@latest add adu2021/cosinegate-dynamic-routing