Non-Parametric Bayesian Inference for Change Point Detection in Neural Spike Trains

We present a model for point processes with gamma distributed increments. We assume a piecewise constant latent process controlling shape and scale of the distribution. For the discrete number of states of the latent process we use a non-parametric assumption by utilizing a Chinese restaurant process (CRP). For the inference of such inhomogeneous gamma processes with an unbounded number of states we do Bayesian inference using Markov Chain Monte Carlo. Finally, we apply the inference algorithm to simulated point processes and to empirical spike train recordings, which inherently possess non-stationary and non-Poissonian behavior.