Paula Harder is about to finish her PhD and is an incoming postdoc at Mila Quebec AI Institute. Fresh from her ICLR 2024 poster session in Vienna, she speaks to us about her paper, which uses deep learning to advance climate science. This article should have been published last month in our BEST OF ICLR special edition. Apologies to awesome Paula and to readers. Computer Vision News Computer Vision News 44 ICLR Poster Presentation Hard-Constrained Deep Learning for Climate Downscaling Climate downscaling, a technique similar to super-resolution in computer vision, has emerged as an innovative new tool for enhancing the resolution of climate data. Paula and the team behind this work have adapted deep learning architectures to augment climate models, enabling more detailed spatial predictions and incorporating constraints to ensure physical plausibility within the neural networks. The motivation for this work stems from the limitations of traditional climate modeling, which often provides coarse long-term predictions over large regions, offering a single temperature value for sizable areas. “Maybe we want to know what’s going to happen here in Vienna,” Paula points out. “That’s where we need higherresolution climate data. The numerical models are just too
RkJQdWJsaXNoZXIy NTc3NzU=