First attention over graph wavelet coefficients: learnable node and scale weights
Pseudomanifold · x · 2026-11-09
Part 7 of the Krishnaswamy Lab thread: standard graph wavelet transforms combine coefficients with fixed, uniform weights; the authors instead apply learnable attention over nodes and scales, interpretable as region and scale importance. They claim this is the first use of attention over graph wavelet coefficients.
More from Research
- Agent-native 3D engine Vork: 2,500-token API, up to 24x faster rendering than Blender — ravisparikh · 2026-11-09
- Predicting AI right-to-left: lab sociology and compute distribution beat tech reasoning — akbirthko · 2026-11-09
- Transluce builds pipeline uncovering thousands of innocuous-looking datasets that induce harmful behavior after finetuning — RexDouglass · 2026-11-09
- Author explains: too expensive to train my own model, so I open-sourced 1.4M sprites — Angaisb_ · 2026-11-09
- jd_pressman on why learning complex specs fails: missing synthetic gradients — jd_pressman · 2026-11-09
- METR says it has 'convincing evidence' AI speeds up research — but the details are redacted — peterwildeford · 2026-11-09