Fused DeepLIFT explained: why attribution hooks are cheap for Borzoi/AlphaGenome but hurt small models
jmschreiber91 · x · 2026-10-26
jmschreiber91 explains the engineering behind "fused DeepLIFT":
- How DeepLIFT works: it overrides gradient ops to incorporate a reference; general-purpose implementations attach "hooks" to model layers so users don't notice the model is modified.
- Cost profile: for large models like Borzoi/AlphaGenome, hook overhead is negligible relative to the forward pass; for small models it matters a lot, since hooks add many small ops that can't run asynchronously, causing extra GPU stalls.
A practical look at productionizing attribution methods for genomics models.
Related event: Fused DeepLIFT explained: bypassing hooks for a 7x speedup(2 posts)→
More from Research
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26
- Ortet evaluates 34 open-weight models on 50 health tasks, finds general leaderboards insufficient — elmanmansimov · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26