Original title: Active machine learning using MCMC Bayesian neural networks
Devesh Jawla, John Kelleher, Maria Chiara Leva · September 2026
Teaching a computer often requires thousands of human-labeled examples, which costs a lot of time and money. Active learning fixes this problem. Instead of labeling everything, the computer points to the specific data points it feels most confused about and asks humans to label only those. To do this, the computer must measure its own doubt. Usually, researchers rely on fast shortcut math to estimate uncertainty, but these shortcuts often produce errors when data is scarce. A slower, more precise math technique called Markov Chain Monte Carlo gives much better estimates, but it is normally too slow for big projects. The researchers realized that active learning keeps data pools tiny, making this slow math fast enough to use. Across eight classification tests, a small network using this exact math beat the shortcut method in six tests, needing only one hundred labeled examples.
Imagine you are trying to guess the weight of a bag of apples by feeling it. You can either take a quick, rough guess or spend time carefully balancing it on a scale. The quick guess is usually good enough, but when the bag is very light, the quick guess can be way off. Here, the quick method is like using a faster but less accurate way to estimate uncertainty, and the careful method is like using a slower but more accurate way. The paper shows that when you have only a few apples (few labels), the careful method gives a much better estimate, helping you decide which apples to check next.
Active learning is used to reduce the cost of labeling data for machine learning. This work shows that using a more accurate Bayesian inference method (MCMC) in a small model can outperform a larger model with a cheaper approximation, achieving high accuracy with very few labels. This means practitioners can get better results with less annotation effort, especially when labels are scarce.
The results are limited to tabular classification datasets with a 100-sample annotation budget and the PowerBALD acquisition function. The MCMC method may not scale to larger datasets or more complex models. The comparison is only against MC Dropout, not other approximate Bayesian methods. The paper does not claim general superiority outside the low-annotation regime.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Active machine learning using MCMC Bayesian neural networks — Devesh Jawla, John Kelleher, Maria Chiara Leva — Open Research Europe — doi:10.12688/openreseurope.24569.1
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.