André Marquand -
Date: - Location: Eurecom
In recent years there has been an enormous interest in applying machine learning methods to clinical neuroscience data to assist diagnosis and predict outcome. For such applications, accurate quantification of uncertainty is crucial and probabilistic methods are a natural choice. In this talk I will present work from our group that uses Gaussian processes for predicting disease state in a range of disorders both in the context of classical supervised learning and in an anomaly detection setting. I will outline methods we have developed to accurately quantify centiles of variation across the population and to scale Gaussian processes up to map spatial correlations across brain imaging datasets containing hundreds of thousands to millions of locations. I will illustrate these methods by highlighting clinical applications across a range of brain disorders.
