Reproduce “Point process models for sequence detection in high-dimensional neural spike trains” on OpenResearch

About this paper

Alex H. Williams, Anthony Degleris, Yixin Wang, Scott W. Linderman

Researchers from Stanford University and Columbia University introduce PP-Seq, a Bayesian point process model for unsupervised detection of sparse neural spike sequences in high-dimensional neural data. The model operates in continuous time, quantifies uncertainty in parameters and predictions, and incorporates learnable time warping, demonstrating enhanced robustness and interpretability on synthetic and experimental datasets.

Key insights

• Formulating neural sequence detection as a continuous-time Bayesian point process provides a statistically robust framework that inherently handles uncertainty and avoids limitations associated with spike discretization.
• An ultra-sparse representation of latent sequence occurrences in continuous time facilitates the natural incorporation of advanced model extensions, such as learnable time warping, to capture complex biological phenomena like variable sequence durations.
• Efficient trans-dimensional inference, achieved through a novel collapsed Gibbs sampler with MCMC improvements, allows complex probabilistic models to be applied effectively to large-scale, high-dimensional neural spike train data.

Summary by alphaXiv.