Active learning using Hamiltonian Monte Carlo

Ward Janssens, Adam Arany, Hannah Rosa Friesacher

Track: Type D (Master/Bachelor Thesis Abstracts)

Keywords: Active learning, Hamiltonian Monte Carlo, Bayesian, Uncertainty Estimation

Abstract: In this project, we use Hamiltonian Monte Carlo (HMC) to perform uncertainty estimation in a regression-based active learning setting. HMC leverages concepts from physics to efficiently sample from the posterior distribution, thereby overcoming the random-walk limitations of classical MCMC algorithms. Using HMC, we estimate the variance of the posterior predictive distribution, which provides a measure of uncertainty. This allows us to quantify the expected reduction in training loss for unlabelled samples if they were queried from the oracle. In active learning, such information is crucial, as the goal is to minimise annotation costs while maximising model improvement. Our method therefore selects the data point with the highest posterior variance for querying. To evaluate its performance, we conduct a simulation study comparing this approach to a baseline that augments the training set with randomly chosen samples. The results show that HMC-based selection achieves a significantly faster reduction in mean squared error compared to random querying. We conclude that HMC offers a viable and effective strategy for active learning in Bayesian models, paving the way toward more scalable applications of Bayesian deep learning.

Serve As Reviewer: ~Adam_Arany1

Submission Number: 14