engleski

A probabilistic deep learning model for predicting missing spatio-temporal data using high-performance GPU computing

We present our winning solution model developed for the recent Data Challenge at the Extreme Value Analysis 2021 conference. The Data Challenge was about predicting deliberately missing parts of spatio-temporal numerical data on historical occurrences of wildfires in the United States. Our approach was developed with only a minimal set of assumptions and no specialist knowledge from the field of extreme value analysis. The conditional missing data importance-weighted autoencoder (CMIWAE) model is based on the variational auto-encoder architecture and is incorporating many of the recent state-of-the-art developments. The distribution of each missing data point is predicted in the absence of the complete dataset. We use a deep neural network inspired by the ResNet and U-Net architecture originally developed for image classification and segmentation tasks. The complete approach is highly parallelizable and greatly benefits from a multi-GPU high-performance computing environment. Authors hypothesize that this model is more general and could be adapted to different domains, including but not limited to: signal processing, super-resolution tasks, and numerical solving of differential equations.

Data reconstruction ; Variational autoencoder ; Convolutional neural network ; Deep learning ; Ensemble ; Extreme Value Analysis Conference challenge ; Wildfires

nije evidentirano