In over-parameterized Bayesian one-hidden-layer networks, class-manifold separation becomes nonmonotonic in temperature and hidden weights develop data-dependent correlations, signatures of feature learning despite Gaussian-process predictions.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cond-mat.dis-nn 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Microscopic and collective signatures of feature learning in neural networks
In over-parameterized Bayesian one-hidden-layer networks, class-manifold separation becomes nonmonotonic in temperature and hidden weights develop data-dependent correlations, signatures of feature learning despite Gaussian-process predictions.