Garanti
Knowing When to Trust Machine Learning Models: A Bayesian Perspective
- Brennon Shanks
Popis
Modern biology increasingly relies on machine learning to model biomolecular interactions, predict molecular properties, and accelerate molecular simulations. These models have remarkable predictive power, but they also make mistakes — and those mistakes are often difficult to identify without quantifying the credibility of the model. Uncertainty estimation is therefore essential for determining when a machine learning model prediction can be trusted and remains one of the central challenges in scientific machine learning. This lecture introduces Bayesian inference as a probabilistic framework for building machine learning models that quantify their own uncertainty, providing a principled measure of confidence in model predictions. Beginning with simple Bayesian parameter estimation, we will develop the fundamental concepts and intuition behind probabilistic inference before extending them to modern machine learning models. We will explore how Bayesian methods enable more reliable scientific inference, improve decision-making, and identify where additional data would be most valuable. Applications will include biomolecular force field design, machine learning interatomic potentials, and other areas of computational biology where understanding model uncertainty is essential.