Machine learning algorithms offer state-of-the-art predictive performance in a variety of domains, but often lack an associated measure of uncertainty regarding its predictions. Split conformal prediction is a leading tool to obtain predictive intervals with virtually no assumptions beyond data exchangeability. This crucial assumption, however, hinders its applicability to many important data, such as time series and spatially dependent processes. In this talk, we will introduce split CP and show how it can be extended to non-exchangeable settings through a small coverage penalty. The proposed framework, based on data decoupling and concentration of measure inequalities, works more generally than traditional split CP, and experiments corroborate our coverage guarantees even under highly dependent data. This is joint work with Roberto Imbuzeiro Oliveira, Thiago Ramos and João Vitor Romano.
Texto informado pelo autor.
* Os participantes dos seminários não poderão acessar às dependências da FGV usando bermuda, chinelos, blusa modelo top ou cropped, minissaia ou camiseta regata. O uso da máscara é facultativo, porém é obrigatória a apresentação do comprovante de vacinação (físico ou digital).
Apoiadores / Parceiros / Patrocinadores
Paulo Orestein - I'm an assistant professor at IMPA. My research is focused on the interplay between statistics, probability, and computation. On the theory side, I am interested in high-dimensional Bayesian models, Monte Carlo methods and robust mean estimation. On the applications side, I have been using machine learning to extend weather forecasts into the subseasonal realm. I enjoy working on both theoretical and applied projects, and find them to often illuminate each other. Before coming to IMPA, I obtained a PhD in Statistics at Stanford University, advised by Persi Diaconis, and received a Masters in Mathematics and BSc in Economics, both from PUC-Rio, in Brazil. I also spent a quarter at Google Research and another at Meta Reality Labs.
a) Opção presencial *
Praia de Botafogo, 190
5o andar, Auditório 537
b) Opção remota (via Zoom)
Meeting ID: 981 5075 6795
Tel: 55 21 3799-5917