Talk by Jorn Diedrichsen (DSAS colloquium)

Title: Statistical approaches for the analysis of brain representations

Abstract: One challenge in the field of neuroscience is to understand how complex pattern of neural activity in the human brain relate to the functional processes that allow for intelligent behavior. Novel recording techniques result in highly multivariate dataset that can measure the activity patterns across many neurons or brain regions for a larger number of conditions. I will focus on two statistical approaches, currently used to analyze the information content of such pattern: In Pattern Component Modeling (PCM) we use a hierarchical Bayesian approach to model the probability distribution of this data directly – allowing us to approximate marginal likelihoods for model comparison. In an alternative approach, Representational Similarity Analysis (RSA), inferences are based on similarity measures between activity patterns as a sufficient statistic. I will show new results that improve inference in the latter approach, linking it theoretically to the Hilbert-Smith Independence Criterion. I will end by showing some of the new challenges in the development of statistical techniques for brain imaging data.