OxCSML seminar this week: Richard Turner

[Hai Dang Dau]

Dear all,

This Friday we welcome Richard Turner from University of Cambridge to give a talk in our OxCSML seminar. Please find details below; looking forward to seeing you there.

Kind regards,

Saif & Hai-Dang

=============

Speaker: Richard Turner, University of Cambridge

Time and date: 14.00 to 15.00 Friday 24 November
Place: Room LG.03 (Small lecture theatre), Department of Statistics, Oxford

Zoom: https://zoom.us/j/98645260498?pwd=S0kyTmJJMkUrT0FiVk5IYjRGYm9jQT09

Title: Neural Processes for Environmental Prediction

Abstract: A host of important problems in the environmental sciences involve processing off-the-grid data in order to make spatio-temporal predictions at arbitrary positions. For example, forecasting air quality requires assimilation of data from air-quality monitors, and forecasting weather requires assimilating data from weather stations, satellites, aircraft and ships. Neural processes are a class of models inspired by stochastic processes like Gaussian Processes, that allow deep learning modules, such as transformers and CNNs, to be leveraged for such tasks. In this talk, I will give an introduction to neural processes and then showcase their use in different applications including active learning for the placement of weather stations, predicting fluid flow, and air quality forecasting. I will end by discussing how machine learning based methods are poised to revolutionise operational weather forecasting. Notably, current efforts have focussed only on replacing the numerical integration system. I will discuss ongoing work in my group where we are starting to develop tools based on neural processes that might instead replace the entire forecasting pipeline including data assimilation, integration, and post-processing.