Interpretability In Atomic-Scale Machine Learning
summary
Claims that atom-scale ML uses representations built from symmetrized correlations of the atom density and a hierarchy of 2-, 3‑body (distance+angle) descriptors expanded in radial functions and spherical harmonics, yielding equivariant features equivalent to modern message‑passing nets. Demonstrated practices: cutoff‑dependent learning curves showing short cutoffs win with small data, multi‑scale weighting improves accuracy, a semi‑local fix for long‑range electrostatics, ablation/knockout tests to find key terms, and recovered alchemical archetypes that cluster by periodic‑table groups.
AI-written summary · about
chapters (9)
- 0:00Introduction and context
- 1:06Atomic scale modeling goals
- 6:12Representation and symmetry
- 10:05Descriptors and frameworks
- 14:18Introspecting models
- 18:44Chemical insights
- 21:23Unsupervised learning
- 24:22Ablation and conclusion
- 25:38Discussion and Q&A
related talks
56mWrap up and reflection talk
01h 15mImproved Machine Learning Algorithm for Predicting Ground State Properties
35mReading the Mind of an LLM
44mQuantum machine learning: Prospects and challenges
34mCausality, robustness, and natural language understanding in ML
59mEquivariant Machine Learning Structured Like Classical Physics
58mStrong Generalization from Small Brains and No Training Data
56mAbstract Domains for Machine Learning Verification