21.07 11:30 - 12:30
USI East Campus, Room D0.02 |
|---|
| |
|---|
Abstract: Strictly proper scoring rules (SPSR), e.g. Energy score and Kernel score, are widely used diagnostics for probabilistic forecasting. As statistical divergences can be derived from SPSRs, they have recently gained popularity for inference in frequentist (as minimum scoring rule estimators) and Bayesian frameworks (as generalised posteriors) when likelihood functions are intractable, replacing negative log-likelihoods with SPSRs. Here, we explore a generalised Bayesian model selection framework for models without tractable likelihoods, using Bayes factors from generalised posteriors. To compute generalised marginal evidence under each model, we propose a path sampling estimator combined with sequential Monte Carlo. We study the procedure's consistency theoretically and through simulations, illustrating its use for selecting challenging biological models without tractable likelihoods.
https://teams.microsoft.com/meet/378844928764790?p=qAEz4YQpgdE6XQLW8d
Host: Prof. D.Sulem | |
|---|
|
|---|
|
|
| | Ritabrata ‘Rito’ Dutta, Reader in the Department of Statistics (Warwick), works on robust simulation-based inference of generative models for probabilistic prediction. His collaboration with ECMWF on projects funded by ECMWF and the Turing Institute focuses on developing robust diagnostic tools to evaluate weather forecast models and provides foundation for training data-driven ML models for weather forecasting and downscaling via diagnostics tools like CRPS, which has become the go-to tool for all existing models. He is one of the pioneers for developing AI models for weather prediction downscaling tasks. He has (co-)led funded research totalling over £1.5M (EPSRC EP/V025899/1, EPSRC EP/T017112/1, NERC NE/T00973X/1), covering optimal lock-down strategies during COVID-19, predicting fish stock movements in the English Channel, or quantifying uncertainties in long-form text generation. 11:30 |
|---|
| |
|---|
|
|
|
|
.jpg?2026-06-10%2010:20:00)
