English

BayesAdapter: Being Bayesian, Inexpensively and Reliably, via Bayesian Fine-tuning

Machine Learning 2022-10-14 v5 Machine Learning

Abstract

Despite their theoretical appealingness, Bayesian neural networks (BNNs) are left behind in real-world adoption, mainly due to persistent concerns on their scalability, accessibility, and reliability. In this work, we develop the BayesAdapter framework to relieve these concerns. In particular, we propose to adapt pre-trained deterministic NNs to be variational BNNs via cost-effective Bayesian fine-tuning. Technically, we develop a modularized implementation for the learning of variational BNNs, and refurbish the generally applicable exemplar reparameterization trick through exemplar parallelization to efficiently reduce the gradient variance in stochastic variational inference. Based on the lightweight Bayesian learning paradigm, we conduct extensive experiments on a variety of benchmarks, and show that our method can consistently induce posteriors with higher quality than competitive baselines, yet significantly reducing training overheads. Code is available at https://github.com/thudzj/ScalableBDL.

Keywords

Cite

@article{arxiv.2010.01979,
  title  = {BayesAdapter: Being Bayesian, Inexpensively and Reliably, via Bayesian Fine-tuning},
  author = {Zhijie Deng and Jun Zhu},
  journal= {arXiv preprint arXiv:2010.01979},
  year   = {2022}
}

Comments

14th Asian Conference on Machine Learning (ACML 2022)

R2 v1 2026-06-23T19:02:35.248Z