Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Susceptibilities for Neural Networks Learning from Physical Data
Rohan Hitchcock, Gary W Delaney, Jonathan H Manton, Richard Scalzo, Jingge Zhu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes susceptibilities from statistical physics as a tool to analyse how a neural network trained on physical data responds to perturbations of physical parameters.
Problem
Neural networks trained on data from physical systems must learn regularities induced by underlying physical laws, and it is unclear how those laws shape the learning process.
Approach
Applies susceptibilities (response of a network to perturbations of a physical parameter of the data-generating system), a concept from statistical physics, as a framework for analysing learning from physical data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| bias training mixture toward small lattice occupation probability h (alpha < 0) (ablation) | 15M transformer on percolating states, 16x16 lattice · percolation, h in [0.35, 0.6] | mean maximum cluster height of generated samples | Improved |
| bias training mixture toward large h (alpha > 0) (ablation) | 15M transformer on percolating states · percolation, h in [0.35, 0.6] | mean maximum cluster height of generated samples | Hurt |
| restrict susceptibility to critical-site losses (masked susceptibility chi_crit) (ablation) | susceptibility analysis · percolation | predictive value for data selection | Improved |
| use full susceptibility chi(h) instead of masked chi_crit(h) to guide data selection (ablation) | susceptibility analysis · percolation | informativeness for data selection | No clear effect |
| compare chi_crit with empirical loss, gradient norm and gradient dot product as diagnostics (ablation) | data-selection diagnostics · percolation | agreement with retraining result | No clear effect |
| restrict susceptibility to attention-head parameters (chi_heads) instead of full-model loss (chi_full) (ablation) | granular-material transformer · granular materials, psi = total collision force | detection of collision-resolution emergence | Improved |
Closest papers by meaning
- Adversarial Attack and Defense for Machine Learning in Statistical Physics (NeurIPS 2026 · similarity 0.87)
- Analyzing Neural Scaling Laws in Two-Layer Networks with Power-Law Data Spectra (ICLR 2025 · similarity 0.82)
- Learning from Integral Losses in Physics Informed Neural Networks (ICML 2024 · similarity 0.82)
- Models of Heavy-Tailed Mechanistic Universality (ICML 2025 · similarity 0.82)
- Critical feature learning in deep neural networks (ICML 2024 · similarity 0.82)
- Stable Port-Hamiltonian Neural Networks (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
