Learning Without Labels - New Insights into Climate and Extremes

Explore how unsupervised learning techniques can revolutionize climate science and weather prediction in this 59-minute conference talk by Maria Molina from the University of Maryland. Discover innovative approaches that move beyond traditional supervised learning methods to uncover hidden patterns in high-dimensional climate data without relying on predefined labels or indices. Learn about cutting-edge applications including knowledge-guided autoencoders that can disentangle distinct Pacific climate modes with different spectral signatures, and custom hyperparameter searches that optimize self-organizing maps to create smooth, interpretable pathways among weather regimes. Understand how these unsupervised learning methods reveal unexpected patterns and mechanisms underlying established climate and weather phenomena while avoiding biases inherent in labeled datasets. Gain insights into how these advances in machine learning can improve our understanding of climate variability, weather extremes, and their implications for prediction, preparedness, and resilience in a changing climate. The presentation was delivered at IPAM's Mathematics and Machine Learning for Earth System Simulation Workshop, demonstrating the intersection of advanced computational methods and climate science for scientific discovery and decision-making.