McGill University researchers have developed a more energy-efficient method of building AI systems that are better at measuring – and indicating – their own uncertainty. This will help users determine when human oversight is needed, when additional data should be collected and when a model is being asked to work beyond the conditions it was trained for, the researchers said.
"Artificial intelligence systems now play a central role in daily life, from medical diagnosis and content moderation to autonomous driving and AI agents that act on our behalf," said Mame Diarra Touré , lead author and PhD Candidate in the Department of Mathematics and Statistics . "As these systems take on more responsibility, they need to become more trustworthy. They should recognize when they are uncertain, rather than giving confident answers in situations where they may be wrong."
The research was supervised by David A. Stephens , Professor in the Department of Mathematics and Statistics.
More reliable estimates
Standard neural networks learn patterns from data and make predictions, but they typically provide a single answer without clearly indicating how confident they are in that response. Bayesian neural networks address this limitation by representing their internal settings as probabilities rather than fixed values, enabling them to estimate uncertainty, particularly when faced with unfamiliar data. However, this capability often requires significant computational and memory resources, making these networks difficult to deploy at the scale of modern AI systems.
The researchers found a way to make Bayesian neural networks substantially more efficient while maintaining strong predictive performance. In one experiment, their approach used about 33 times fewer parameters than a commonly used method for estimating uncertainty in AI systems.
The results suggest that reliable, uncertainty-aware AI can be made practical even for the large and complex systems in use today, Touré said.
The researchers are now exploring ways to automate the process of identifying which parts of a neural network are most important for a given task. This could help the approach work more effectively across different kinds of data and AI tools, they said.
About this study
" Singular Bayesian Neural Networks ," by Mame Diarra Touré and David A. Stephens, was presented at the Forty-Third International Conference on Machine Learning (ICML 2026).
NB: Machine learning research follows a different publication model from many other scientific fields. The top peer-reviewed conferences are often the main archival publication venues, rather than a preliminary step before journal submission.
ICML papers are reviewed through a double-blind peer-review process, and accepted papers are published through the Proceedings of Machine Learning Research.
