Data-Driven Priors for Trustworthy Machine Learning

Machine learning models, while effective in controlled environments, can fail catastrophically when exposed to unexpected conditions upon deployment. This lack of robustness, well-documented even in state-of-the-art models, can lead to severe harm in high-stakes, safety-critical application domains such as healthcare. This shortcoming raises a central question: How can we develop machine learning models we can trust?

In this talk, I will approach this question from a probabilistic perspective and address deficiencies in the trustworthiness of neural network models using Bayesian principles. Specifically, I will show how to improve the reliability and fairness of neural networks with data-driven domain-informed prior distributions over model parameters. To do so, I will first demonstrate how to perform training in neural networks with such priors using a simple variational optimization objective with a regularizer that reflects the constraints implicitly encoded in the prior. This regularizer is mathematically simple, easy to implement, and can be used as a drop-in replacement for existing regularizers when performing supervised learning in neural networks of any size. I will then show how to construct and use domain-informed data-driven priors to improve uncertainty quantification and group robustness in neural network models for selected application domains.