The Netherlands Organisation for Scientific Research (NWO) has awarded a Physical Sciences TOP grant 1 for curiosity driven research to Peter Grünwald of CWI. Module 1 is a grant for senior researchers with a proven track record of significant research results in the last ten years. These are researchers who because of their past performance are regarded as being among the best in their field. Grunwald will receive funding for the appointment of a maximum of three research positions (PhD students and/or postdocs), plus a maximum of 25,000 euro for additional resources for the proposed research.
Grünwald has received the TOP grant 1 for his project 'Safe Bayesian Inference: A Theory of Misspecification based on Statistical Learning'. In this project, Grünwald will further develop the Safe Bayesian Method and related methods, which have good predictive behavior even if the model is wrong-yet-useful. These methods are unique in that they combine ideas from modern machine learning theory with those of classical statistics. Much of modern statistics and machine learning is based on Bayesian methods. Nowadays, these are often applied with highly complex models, involving 10000s of free parameters - such models are often highly useful, yet usually not fully correct: for example, when fitting a curve, one assumes that noise is normally distributed where in reality, the noise is more complicated.
Grünwald and his group have recently shown that in such situations, Bayesian methods can fail dramatically, leading to very bad predictions of future data. For example, when predicting future temperatures based on a database of daily temperatures in Seattle, the Safe Bayesian method vastly outperforms standard Bayesian methods such as the Bayesian Lasso.