SA-SGLD adapts SGLD stepsizes via gradient-norm-based time rescaling to sample BNN posteriors more accurately than standard SGLD on toy examples and image classification tasks without introducing bias.
Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Adaptive Stepsizing for Stochastic Gradient Langevin Dynamics in Bayesian Neural Networks
SA-SGLD adapts SGLD stepsizes via gradient-norm-based time rescaling to sample BNN posteriors more accurately than standard SGLD on toy examples and image classification tasks without introducing bias.