This past Tuesday, October 15th, Inria researchers had the chance to hear from Douglas Brum, a visiting student from Federal Fluminense University. Douglas is working on some fascinating research in collaboration with Mariza Ferro and Luan Teylo. His work focuses on using deep learning models to predict extreme weather events, particularly heavy rainfall in areas like Rio de Janeiro.
Douglas’s presentation tackled a problem that feels more pressing every year: extreme weather events are becoming more frequent and more intense, largely due to human impact on the planet’s natural systems. Pollution, deforestation, and other activities have led to a worrying increase in severe precipitation. Cities like Rio, with their complex landscapes, face added challenges in accurately predicting these events.
At the heart of his work is the PredRNN-V2 model, a stacked recurrent neural network designed to improve the accuracy of weather predictions. But as anyone working in AI knows, the computational cost of deep learning models can be sky-high. To address this, Douglas is exploring a multi-scale recurrent neural network model that offers a more efficient alternative, aiming to keep the same level of prediction accuracy while reducing the strain on resources.
What’s particularly exciting is that the model is expected to be tested with real-world data from Rio in the near future, which could bring a lot of value to local forecasting systems.
