Gaussian process based nonlinear latent structure discovery in multivariate spike train data.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Data Analysis

Background:

• Extracting low-dimensional latent structure from multi-neuron spike train data is a significant challenge.
• Existing methods often rely on linear assumptions for latent dynamics or spike rate mappings.
• Neural data, such as hippocampal place cell codes, can exhibit high linear dimensionality.

Purpose of the Study:

• To propose a doubly nonlinear latent variable model capable of identifying low-dimensional structure in high-dimensional spike train data.
• To introduce a novel and efficient approximate inference method for model learning.
• To demonstrate the model's effectiveness in analyzing real neural data.

Main Methods:

• The Poisson Gaussian-Process Latent Variable Model (P-GPLVM) combines Poisson spiking observations with two Gaussian processes for temporal dynamics and nonlinear tuning curves.
• A decoupled Laplace approximation method is introduced for fast and accurate approximate inference, optimizing latent paths while marginalizing tuning curves.
• The model is applied to spike train data from hippocampal place cells.

Main Results:

• The P-GPLVM successfully identifies low-dimensional latent structure even in data with high linear dimensionality.
• The decoupled Laplace approximation demonstrates superior speed and accuracy compared to previous inference methods.
• The model achieves competitive or superior performance against established methods, including Variational Auto-Encoder (VAE) based approaches.

Conclusions:

• The P-GPLVM offers a powerful new tool for uncovering latent structure in complex neural population activity.
• The decoupled Laplace approximation provides an efficient inference solution for nonlinear latent variable models.
• This approach advances the analysis of neural coding and brain function by revealing underlying low-dimensional dynamics.