FlexGlue author shares training details: entmax loss, MLP refinement, per-cell query sampling
ducha_aiki · x · 2026-11-05
The FlexGlue author shared paper detail screenshots highlighting design choices:
- Sub-pixel readout: windowed softmax replaced by 1.5-entmax with a sparse entmax loss (exact zeros for unlikely cells), supervised via cross-entropy against bilinear ground-truth weights over four neighboring cells — log-space interpolation is crucial or mass collapses onto the nearest cell.
- Local refinement: coarse readout already puts >80% of queries within 2px; a small MLP predicts ±1-cell sub-cell offsets from interpolated layer tokens, needing only two interpolations per query and scaling linearly; gradients through the coarse position are stopped, refinement shifts matches only 0.6px on average.
- Supervision: one query sampled per 16×16 cell with random offsets (no keypoint detector dependency, queries swappable at test time); balanced BCE for matchability and covisibility, total loss 10Lcoarse + Lfine + Lmatch with deep supervision.
Related event: LightGlue Team Releases FlexGlue, Topping IMC2025 Benchmarks(5 posts)→
More from Research
- RL theory monograph by Chen, Chi, Fan et al: Statistical Foundations of RL, a non-asymptotic view — burny_tech · 2026-11-05
- Weight-space programming reproduced: GPT-6.1 beats PPO baseline 21x, no open-weight model passes it — burny_tech · 2026-11-05
- 8% of Nature peer reviews since June 2025 are fully AI-generated, analysis finds — TuhinChakr · 2026-11-05
- HandsOn benchmark: best dexterous manipulation policy manages only 29.1% success on Core-50 — jiqizhixin · 2026-11-05
- David Baker gives Royal Society Croonian Lecture on computational protein design — mmbronstein · 2026-11-05
- Paper: rewriting reasoning histories into compact records cuts agent token costs — Anton Hristov · 2026-11-05