The EnviBayes section is pleased to announce the resumption of its monthly webinars. We have a great lineup for you this fall.
The first speaker will be Andrew Zammit Mangion from University of Wollongong. The webinar will take place on Tuesday, September 20 at 4:30pm Eastern Time in the US (Tuesday, September 21 at 6:30am local Wollongong time).
Registration is required using the following link: https://zoom.us/meeting/register/tJUofu-sqzopHdUqlTonK5HWFlyJNrb-mZW-
Title: Deep learning for facilitating parameter estimation in spatial and spatio-temporal models
Parameter estimation is often the computational bottleneck in an environmental data analysis involving statistical spatial or spatio-temporal models. This talk will present two cases in which deep learning can help alleviate this computational burden. In the first part of the talk, I show the utility of a convolutional neural network (CNN) for estimating dynamical parameters in a statistical spatio-temporal process model. Once the CNN is fitted, probabilistic forecasting in a setting where dynamics are changing rapidly both in space and in time can be done extremely quickly and efficiently online. In the second part of the talk, I show the utility of so-called permutation-invariant neural networks for estimating parameters within intractable physical or statistical models from exchangeable replicates. In experiments involving spatial models of extremes, I show that these “neural Bayes estimators” considerably outperform other neural-network-based estimators that do not account for replication appropriately in their network design, and that they are highly competitive and much faster than traditional likelihood-based estimators. The talk will also illustrate two environmental case studies where the deep learning estimators are put to good use. The first part of the talk is joint work with Christopher Wikle (University of Missouri) while the second part is joint work with Matthew Sainsbury-Dale (University of Wollongong), and Raphael Huser (KAUST).