Program Lead NOAA Fisheries Santa Cruz, California, United States
Machine learning (ML) models often have impressive success in tasks such as classification, computer vision, etc. However, we still don’t have a clear understanding of why these models work, leading occasionally to major surprises when they fail. There is therefore a clear need to connect ML to fundamental theory. Here I describe how recurrent neural networks, a popular ML architecture, share a common theoretical foundation with Empirical Dynamic Modeling based on Takens theorem of time-delay embedding and suggest how we can use this connection to build ecologically reasonable constraints into both approaches.
