SIParCS 2023 - Belen Saavedra

Using Machine Learning Uncertainty Estimates to Aid Scientific Analysis

Recorded Talk

Both weather forecasters and researchers want to know when they should trust the guidance from their models as well as when not to. Being able to understand the sources of uncertainty in their guidance can enable them to convey to decision makers when to wait for further updates versus taking immediate protective actions. However, traditional machine learning models can only provide limited estimates of uncertainty within the realm of their training experience, so they tend to be overconfident in their predictions, especially when being applied to unseen data. A new class of machine learning models called evidential models can estimate the total uncertainty of a single prediction by making strong prior assumptions about how the data and model are expected to behave. In this project, the intern will work with CISL’s machine learning group to develop new ways to analyze and visualize uncertainty estimates and explanations of the uncertainty from these evidential models for multiple weather forecasting use cases, including winter precipitation type and severe storm hazard (e.g., tornadoes, hail) prediction.

Mentors: David John Gagne, Charlie Becker, Gabrielle Gantos, John Schreck

Slides and poster