This is a project that I completed during my first semester as a graduate student at CMU for a ML and AI class (24-787). The goal of the project is to predict and understand trends in global indicators (such as GDP or food production). To accomplish this, a variety of approaches are investigated. Most notably, I attempt to implement two sequence to sequence algorithms inspired by machine translation applications: an encoder decoder LSTM architecture and a WaveNet inspired time series CNN. I'd like to thank Joseph Eddy for his fantastic introductions to time series forecasting with sequence to sequence models (https://github.com/JEddy92/TimeSeries_Seq2Seq/blob/master/notebooks/TS_Seq2Seq_Intro.ipynb). I relied on his code to get these architectures off the ground.
Also, I was very interested in exploring machine learning approaches that could lend themselves towards more interpretable models. I played around with Dynamic Mode Decomposition (using this fantastic library https://github.com/mathLab/PyDMD#references) and my results were pretty good! I am deeply intrigued by potential applications of the work of Brunton et al in identifying sparse dynamics for high dimensional nonlinear systems (https://arxiv.org/abs/1509.03580). I think Brunton's work and other ongoing work in DMD and Koopman theory represent potentially revolutionary applications of ML; by allowing rapid discovery of governing dynamics from data alone, these approaches could really change research in fields as diverse as physics, economics, sociology, and control engineering (essentially any discipline that seeks to model a system's behavior).