Yasir Alsugair

Artificial Intelligence Specialization · Cohort 2026

Toward Scalable Bayesian Uncertainty for Machine Learning with the Ray Tracing Sampler

Machine learning models usually give one answer with no sense of how confident they should be. Bayesian methods fix this by averaging over many plausible versions of the model, but the standard sampling algorithms misbehave when data is fed in small batches, which is how modern machine learning is trained. This project evaluates the Ray Tracing Sampler (Behroozi 2025), a new method built for exactly that setting. It explores the space of possible models the way a light ray bends through glass: it always moves at the same speed, so noisy training signals can steer it but cannot make it overshoot. We first tested it on small neural networks where the exactly correct answer is known, and it reproduced that answer on MNIST. We then applied it to a real astronomy task: predicting a chemical property for about 126,000 stars from Gaia telescope data, with an error bar for every star. Compared with standard training and two popular alternatives on the same likelihood, it matched their accuracy and gave the most trustworthy error bars.

Read the final report (PDF)