New preprint by brilliant @JosueNassar in my group. A new state-space model with multi-scale hierarchical dynamics prior. arxiv.org/abs/1811.12386 Collaboration with @scott_linderman and Monica Bugallo #sciencestorm #neuroscience #Bayesian #MachineLearning 1/6

@JosueNassar @scott_linderman 2/6 Learning dynamics is key to understanding neural computation, but they can be too complex to understand. Can we view it at different levels of granularity? E.g. Lorenz system approximated with 2 linear dynamical systems?

@JosueNassar @scott_linderman 3/6 Or with 4 linear dynamical systems (LDS)? Both of these are from a single model.

@JosueNassar @scott_linderman 4/6 We propose a generative model, TrSLDS (Tree-structured recurrent switching LDS), extending @scott_linderman’s rSLDS (AISTATS 2017) with fully Bayesian sampling using Polya-Gamma augmentation and tree-structured stick breaking.

@JosueNassar @scott_linderman 5/6 Just from the spike trains, we can recover decision making dynamics from the XJ Wang’s 2002 spiking neural network model. (in another paper with @yuanz271 catniplab.github.io/publications/N…)

@JosueNassar @scott_linderman @yuanz271 6/6 It can also recover limit cycles from population V1 spike trains! Looking forward to applying to more neural data!