A min-min objective that jointly learns an energy model and its log-partition provably recovers maximum likelihood in function space and gives an MCMC-free training scheme for combinatorial output spaces.
Title resolution pending
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
-
Joint Learning of Energy-based Models and their Partition Function
A min-min objective that jointly learns an energy model and its log-partition provably recovers maximum likelihood in function space and gives an MCMC-free training scheme for combinatorial output spaces.