Please join us for CIRES HDR, Samual MacDonald’s, PhD confirmation milestone seminar.

Bayesian uncertainty and data set shifts

Speaker: Samual MacDonald

Abstract: Bayesian uncertainty in machine learning is critical to (a) proxy reliability and (b) control the sensitivity/specificity of products in production. Further, Bayesian learning improves generalisation, which is crucial, since ML systems deployed in the real-world are exposed to data set shift, or ‘distribution shift’. Distribution shift renders assumptions inherent to ML unfaithful, which commonly leads to overconfidence. Consequently, ML is rarely trusted, or adopted, by the people responsible for decisions in healthcare.

Healthcare applications relating to cancer are particularly challenged by distribution shift due to biological and data curation factors confounding. Following all of that, this seminar traces a few Bayesian learning techniques both in theory and practice to oncology. Results from (traditional bulk-scaled) transcriptomics support claims that approximate Bayesian extensions to neural networks remedy the   ‘shift-induced overconfidence’. Finally, (modern single-cell-scaled) spatial transcriptomics presents statistical challenges directing (potential) research pathways forward.