A new research paper published on arXiv demonstrates that discrete distributions can be learned even when sampled from metastable states. The study shows that for distributions with strong metastability conditions, the single-variable conditional probabilities closely approximate those of the true stationary distribution. This allows for effective learning of the true model using conditional-likelihood estimators, even when the samples are restricted. The findings are extended to Ising models, providing parameter and structure learning guarantees, and are numerically validated on higher-alphabet spin glass models. AI
IMPACT This research could improve the robustness of machine learning models trained on imperfect or limited datasets.
RANK_REASON Academic paper published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
