DeepLIFT attribution extended to Transformers in tangermeme v1.5.0
jmschreiber91 · x · 2026-10-18
tangermeme, an interpretability toolkit for genomic sequences, merged a PR extending DeepLIFT/SHAP to Transformer architectures (available in v1.5.0):
- Closed-form DeepLIFT rules for LayerNorm, RMSNorm, Softmax, and Bilinear products (matmul/einsum)
- A hook factory for local-IG (local Integrated Gradients) DeepLIFT multipliers
- A routine for locally dinucleotide-shuffled sequences
The maintainer notes the closed-form LayerNorm solution already works with Cherimoya and is fast, and should in principle transfer to Transformer models like Borzoi — enabling faster attribution for long-sequence genomic models.
More from Research
- Paper argues mechanistic interpretability needs a rigorous "logic of inference" — burny_tech · 2026-10-18
- NeurIPS 2026 Journal Track accepts 88 papers from TMLR, JMLR, AoS — NeurIPSConf · 2026-10-18
- RCT: AI-generated layperson summaries of medical findings improve visit quality and shorten consultations — EricTopol · 2026-10-18
- Progress on mapping jev's j-space — sloppenheimer · 2026-10-18
- Hand-probing tiny model neurons: gate is swappable, down steers it into an owl — maxsloef · 2026-10-18
- MIT paper builds attention-based foundation model for quantum states, 18 training points suffice — jwt0625 · 2026-10-18