Events

When each observation is a high-dimensional vector, network, or image, how can we tell whether two populations differ? This can be genuinely difficult: the geometry of the space can work against classical summaries, causing even sensible, well-calibrated methods to miss important distributional differences. This talk presents a different way of reading high-dimensional data by using patterns created by the curse itself as statistical information. Rather than treating these geometric effects only as nuisances, we exploit the systematic structure they create among observations. Starting with two-sample testing, the talk shows how this viewpoint develops into a broader program for statistical inference across different data structures and questions, and concludes with its application to generative-model evaluation, a growing challenge in modern AI.