Reproduce “Statistical mechanics of low-rank tensor decomposition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jonathan Kadmon, Surya Ganguli
Researchers at Stanford University and Google Brain developed Bayes-optimal Approximate Message Passing (AMP) algorithms for generic, arbitrarily shaped low-rank tensor decomposition with mode-dependent priors. The work precisely characterizes inference performance using dynamic mean-field theory, identifying distinct easy, hard, and impossible recovery regimes, and demonstrates superior noise robustness compared to the commonly used Alternating Least Squares (ALS) algorithm.
Key insights
- The existence of an 'easy' inference regime for generic tensors does not necessarily require all modes to have non-zero mean priors, a finding that differs from prior work on symmetric tensors.
- The shape of the tensor (relative dimensions of its modes) significantly influences the width of the 'hard' inference regime, which is maximized for cubic tensors, implying unique properties for these structures.
- AMP algorithms, based on Bayesian inference and rigorously analyzed by dynamic mean-field theory, offer a principled approach to characterize Bayes-optimal performance, unlike heuristic methods such as ALS.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.