Leveraging Interpretable Neural Networks for Scientific Discovery

Abstract:

The past few years have shown huge increases in the use of machine learning for geophysics research. While neural networks are often viewed as black boxes, methods now exist to help interpret the decision making process of the network, helping to optimize network design as well as gain trust in the predictions. Here, we discuss how these tools can also be used to extract new, scientific knowledge, from the data. Examples are provided across a range of examples, including climate change, climate variability, and subseasonal prediction. We further discuss new methods from the computer science community to quantify the uncertainty in the explanation provided by the network interpretability method.