A Poisson-based Parallel State Space Model For Inferring Interregional Neural Communication
Author(s)
Liu, Christopher
Advisor(s)
Hoffman, Courtney
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This work presents a scalable and biologically realistic framework for modeling interregional neural communication using a Poisson-based switching linear dynamical system (SLDS). By replacing Gaussian assumptions with a Poisson observation model, it captures the discrete, burst-like nature of neural spike trains. The inference procedure is reformulated for parallel execution, batching both discrete state estimation and continuous latent optimization using GPU-accelerated operations within a variational Laplace EM framework. Results show that the approach matches the accuracy of sequential methods while achieving nearly a 2× speedup, enabling efficient analysis of large-scale neural datasets.
Sponsor
Date
2026-05
Extent
Resource Type
Text
Resource Subtype
Undergraduate Thesis