TF-MoDISco optimization merged: 5-15x speedup, 2-3x memory savings, identical output
anshulkundaje · x · 2026-10-17
- The Kundaje Lab merged an auto-optimized version of TF-MoDISco delivering 5-15x speedups and 2-3x memory savings with exactly identical output.
- Validated across 4,972 jobs spanning ISM, DeepLIFT, Integrated Gradients and gradients crossed with 60 models, plus malformed inputs; speedups grow with task complexity.
- TF-MoDISco discovers sequence motifs from model importance scores (e.g., BPNet with DeepLIFT/SHAP), a staple of genomic deep learning interpretability. Code is open source on GitHub (kundajelab/tfmodisco, MIT license).
More from Research
- OpenAI researcher: hashing an adversary-chosen timestamp doesn't stop rigging — csvoss · 2026-10-17
- Is SWA the frontier secret sauce? DeepSeek and Kimi chose the hard independent route — burny_tech · 2026-10-17
- Evolving competing neural cellular automata show lifelike self-organizing behavior — burny_tech · 2026-10-17
- Vals AI moves recursive self-improvement forecast to July 2027 as Opus 5.5 tops RSI index — 141_1337 · 2026-10-17
- HF Transformers adds activation checkpoint offloading to CPU to free GPU memory — StasBekman · 2026-10-17
- S4VY: Segment Anything in 4D, with consistent identity across space and time — realChrisChoy · 2026-10-17