Tag: interpretability

All the writing and notes filed under "interpretability" — 1 entry.

• First-Class Per-Site Importance in Materials: Attention-MIL on a Frozen Pretrained Backbone→

  19 May 2026

  · 66 min ·

  • #machine-learning
  • #materials-science
  • #foundation-models
  • #attention-mil
  • +4
  • +2

  A gated attention-MIL pool on top of a frozen pretrained universal MLP (UMA-s-1p2) for materials property prediction. Three-regime wedge characterizing when MIL beats mean-pool vs supervised oracle; literature-curated per-site importance eval with top-3 recall 71.4% [61.7, 80.6] vs random null 10.0%; AOPC faithfulness win over post-hoc attribution comparators; K-fold × ensemble conformal recipe with verified α-sweep coverage. Manuscript in preparation for ICML 2027; arXiv pre-print to follow.