Dr Vincent Fortuin

Vincent Fortuin

Research interestsMachine learning using deep neural networks is omnipresent these days, even though those models can often be overconfident in their predictions and fail in unexpected ways. Bayesian deep learning tries to remedy this by using insights from probabilistic inference, thus yielding more robust and reliable models that can assess their own uncertainty.

In his PhD, Dr Fortuin studied the choice of prior distributions for Bayesian deep learning models. While this problem is often overlooked in the literature, it is crucial for successful Bayesian inference in general. Specifically, Dr Fortuin showed that the choice of priors can have a dramatic impact on the predictive performance of these models. Moreover, he demonstrated how better priors can be identified, thereby enabling entirely new application areas of Bayesian deep learning and potentially making the models more interpretable and trustworthy.

During his Fellowship at St John's, Dr Fortuin is using these ideas to work towards machine learning algorithms for critical applications that require fewer training data, while offering principled model selection and calibrated estimates of their predictive uncertainty.