All Climate Events

Monday, October 28, 2019. 12:00PM. Webinar: Tools for Interpreting how and what neural networks learn, and their applications for climate and weather. Imme Ebert-Uphoff, CIRA; Elizabeth Barnes and Ben Toms, Colorado State University. Sponsored by STAR Science Seminar Series. More information here. 

Artificial neural networks (ANNs) have emerged as an important tool for many environmental science applications. However, ANNs are not naturally transparent and are thus often used as a black box, i.e. without detailed understanding of their reasoning. Fortunately, new tools for the interpretation of ANN models are becoming available from the field of explainable AI. Such tools can provide great benefits for earth science researchers. In this tutorial we first provide a general overview, including methods for both ANN visualization and ANN attribution. Then we focus on one method in detail, namely layer-wise relevance propagation (LRP; sometimes known as Deep Taylor decomposition), and show how it can be used to identify the specific elements of the input that were most important for the ANN's prediction. Thus, this method helps "open the black box" and attribute specific predictions to specific predictands. We find LRP methods to be particularly useful, yet few in the earth science community seem to have discovered them. We demonstrate the use of LRP methods for a variety of applications related to weather and climate, and show their use for tasks ranging from debugging and designing ANN networks to gaining new scientific insights for atmospheric science applications.