McGill University researchers have developed a method for building Bayesian neural networks that can measure their own uncertainty using roughly 33 times fewer parameters than a standard approach, according to a paper presented at the Forty-Third International Conference on Machine Learning (ICML 2026). The technique makes uncertainty-aware AI systems more practical to deploy at scale, which matters for IT teams responsible for building and maintaining systems that need to know when to flag a response for human review.
Bayesian neural networks differ from standard neural networks in how they represent internal settings. Instead of fixed values, they use probabilities, which lets them estimate uncertainty when faced with unfamiliar data. That capability has traditionally required heavy computational and memory resources, making these networks difficult to run within the constraints of modern production systems.
"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.
How the efficiency gain works
Standard neural networks learn patterns from data and make predictions, but they typically return a single answer without indicating how confident they are in that response. Bayesian networks address this limitation by treating their internal settings as probability distributions, which enables uncertainty estimates when the model encounters out-of-distribution input.
The McGill team found a way to preserve that uncertainty estimation capability while cutting the parameter count substantially. In one experiment, their approach used about 33 times fewer parameters than a commonly used method for estimating uncertainty in AI systems, while maintaining strong predictive performance.
Touré said the results suggest reliable, uncertainty-aware AI can be made practical even for the large and complex systems in use today. For engineering teams, this could mean deploying models that flag when they are working beyond their training conditions or when additional data collection is needed, without the infrastructure overhead that Bayesian methods typically require.
What the researchers are working on next
The team is 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.
ICML papers are reviewed through a double-blind peer-review process, and accepted papers are published through the Proceedings of Machine Learning Research.
Why this matters for IT and development
For teams responsible for building and maintaining AI systems, the practical takeaway is that uncertainty estimation no longer has to be treated as a research luxury. The parameter reduction means models can communicate when they are likely wrong without requiring the memory budgets that previously made Bayesian approaches prohibitive. That opens the door to more reliable automation in areas like content moderation, autonomous decision-making, and AI agents that act on user behalf - where knowing what you don't know is often more operationally valuable than the prediction itself.
For developers and data scientists looking to apply these techniques, the underlying mathematics of Bayesian neural networks and uncertainty quantification is a core part of modern machine learning practice. See the AI Learning Path for Data Scientists for structured coverage of these topics, or browse AI for IT & Development for training aimed at engineering roles.
Your membership also unlocks:
