When Deep Learning Falls Short: Neural Networks for Small Datasets in Climate Science
Abstract:
Deep learning thrives on large volumes of data, yet some of the most critical scientific questions of our time must be answered with remarkably small datasets. Global average temperatures from thermometer records span roughly 150 years, reliable data on climate-driven migrations cover only the last few decades, and Antarctic ice cores yield about a thousand data points spread over 800,000 years. In all these settings, conventional deep learning strategies are simply not viable. Neural network techniques, however, retain a distinctive advantage: their ability to uncover nonlinear relationships between variables, which traditional dynamical models may fail to capture. This seminar presents a compact review of climate applications where a neural network tool specifically designed for short datasets has provided new insights, from global warming attribution and the role of human versus natural forcings, to the analysis of climate-migration links in the Mediterranean region. The talk will highlight how these data-driven analyses complement and, in several cases, strengthen the conclusions of physics-based climate models, offering an independent line of evidence on the causes and impacts of climate change. A possible collaborative study on nonlinear causality relationships in a paleoclimatic context will also be proposed.
Short Bio:
Antonello Pasini is Senior Researcher and Climate Physicist at the Institute of Atmospheric Pollution Research of the Italian National Research Council (CNR-IIA) in Rome, and teaches Physics of Climate at the University of Roma Tre. His research focuses on developing and applying mathematical models, in particular artificial neural networks, to identify the causes of climate change at global and regional scales and to study its impacts on territories, ecosystems and societies. He has published extensively in international journals, including Scientific Reports (Nature group), and edited the volume "Artificial Intelligence Methods in the Environmental Sciences" (Springer). He has also served as Vice-President of the Italian Society for Climate Sciences. A committed science communicator, he authored several popular books, among which "Effetto serra, effetto guerra" (ChiareLettere, 2017, with G. Mastrojeni), "L'equazione dei disastri" (Codice, 2020) and "La sfida climatica" (Codice). His blog "Il Kyoto fisso", hosted by Le Scienze (the Italian edition of Scientific American), was awarded the Italian National Prize for Science Communication in 2016.