Last-layer empirical Bayes trains a normalizing-flow prior on the final layer by maximizing expected log-likelihood, and matches but does not beat existing uncertainty quantification methods.
Stochastic gradient hamiltonian monte carlo
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Last Layer Empirical Bayes
Last-layer empirical Bayes trains a normalizing-flow prior on the final layer by maximizing expected log-likelihood, and matches but does not beat existing uncertainty quantification methods.