…on the development and use of Structured Variational Autoencoders to better understand behavioral (and other classses) of data, posted today to the arXiv. The paper is super-cool – it describes a method to learn representations of complex time-series data (which are often non-linear) that are well-suited to structured models (using a Hidden Markov Model, for example). The approach could be used to solve many practical problems in modeling behavioral data, and we are really excited to try it on our 3D datasets. Congrats to Matt and Alex (and of course Ryan Adams) on their great work!