Northstar:“Develop iterative, meaningful benchmarks for AI researchers that enable substantial progress on problems related to climate change as well as impactful AI methodology.”
Summary: Learning from Observational, Multimodal, Multiscale, Spatiotemporal (OMMS) data sources are critical for researchers and practitioners working on problems related to climate change. AI methods for handling these types of data – and the many associated problems – remain largely undeveloped, and researchers/practitioners working on these problems often oversimplify their data sources or cobble together complicated systems from existing tools. To incentivize and encourage the AI community to develop methodology for learning from OMMS data sources, we will launch a benchmark and associated competition/workshop focused on a particular subset of meaningful, yet accessible, prediction problems from OMMS sources.
The first meaningful OMMS benchmark developed by BAIR-Meta will focus on predicting the amount of snowfall in mountainous regions and then further determining how much of this snow will become usable water. This problem, called The Fate of Snow, is a longstanding unsolved problem that is of critical importance to water resource managers, hydrologists, and climate scientists. Making progress on this problem will involve leveraging information from 30+ multimodal data sources and will necessitate innovation in areas including:
1) Learning from sparse, multiscale and multimodal data, e.g. leveraging sparse supervision of dense predictions across different data sources at different locations and time. Typical approaches to multimodal prediction assume dense, input/output data, where for example, each modality has coincident data of similar resolution. This challenge will explicitly rely on sparse datasets of multiple spatial and temporal resolutions, where the model will need to be able to learn how to incorporate each modality into a final prediction.
2) Explainable, probabilistic predictions, e.g. grounding predictions to quantities within physical models and incorporating domain knowledge and/or physical models into the learning process in order to estimate associated uncertainties. Given the sparse, and sometimes contradictory, labeled data, this challenge will encourage participants to incorporate the associated physical models and domain knowledge as inductive biases throughout the learning process, in order to produce probabilistic predictions that have a clear association with underlying physical quantities.
BAIR Researchers:
- Colorado Reed
- Ritwik Gupta
- Prof. Trevor Darrell
Meta AI Researchers:
- Sal Candido
- Matt Uyttendaele
- Joe Spisak
- Ammar Rizvi