Timothy DelSole

Professor in the Department of Atmospheric, Oceanic, and Earth Sciences at George Mason University

Title- Machine Learning Problems in Sub-Seasonal Prediction

Abstract- In this talk, I will discuss a few problems in sub-seasonal prediction that provide rich opportunities for machine learning.  Sub-seasonal prediction refers to weather predictions over time scales of 2-8 weeks and is an emerging frontier in operational forecasting.  First, I illustrate some common pitfalls with applying machine learning algorithms to observational data.  Hopefully, this example will illustrate why many climate scientists are generally skeptical of "big data" methods and serve to clarify an approach to deriving results that are more convincing to climate scientists.  Next, I frame data science as a problem of determining parameters in a "sparse" framework.  Under this view, the major problem in combining data science and climate science is defining the correct sparse framework.  For sub-seasonal prediction, I suggest that one part of this framework is the principle that small-scale structure has less predictive value than large-scale structure.  I then show a relatively new approach to incorporating this principle in a regression framework.

 

Timothy DelSole