Conditionally-Conjugate Gaussian Process Factor Analysis for Spike Count Data via Data Augmentation

Yididiya Y. Nadew, Xuhui Fan, Christopher J. Quinn

Introduction

In modern neuroscience, researchers frequently record the simultaneous activity of hundreds of neurons. To extract meaningful signals from this high-dimensional data, Gaussian Process Factor Analysis (GPFA) is often employed to identify low-dimensional, smooth "latent trajectories" that drive neural populations. However, standard GPFA assumes continuous Gaussian activity, whereas neurons fire in discrete spikes. Modeling these counts using Poisson or Negative Binomial distributions typically breaks mathematical conjugacy, making Bayesian inference computationally expensive and prone to overfitting. This paper introduces ccGPFA, a framework that restores conjugacy through data augmentation, achieving significant improvements in both accuracy and speed.

Methods

The core innovation of ccGPFA is a data augmentation strategy that transforms an intractable inference problem into one with closed-form updates. By introducing auxiliary variables—specifically Polya-gamma, Gamma, and Polya-Inverse Gamma (P-IG) variables—the authors effectively "mask" the non-conjugacy of the spike count likelihood. When conditioned on these variables, the model takes a Gaussian form, enabling Variational EM with simple, analytic coordinate ascent updates.

The probabilistic graphical model of ccGPFA illustrating the hierarchical structure and latent variables.

Furthermore, ccGPFA incorporates sparse Gaussian Processes using inducing points and natural gradients. This allows the model to scale to long, continuous recordings that would otherwise exceed standard memory limits. Unlike previous approaches, ccGPFA also treats the dispersion parameter as a random variable, allowing for a fully Bayesian treatment of over-dispersed neural data.

Datasets and Experimental Setup

The model was evaluated using real-world neural data from the Allen Brain Observatory (Visual Coding Neuropixels Dataset). The researchers analyzed recordings from 176 neurons in the mouse primary visual cortex (V1) during exposure to "drifting gratings" (moving bar patterns). The dataset comprised 75 trials, each 2 seconds long and binned into 15ms intervals. The model was trained on 50 trials and evaluated on the remaining 25.

Key Results

ccGPFA was compared against three major baselines: standard GPFA, Polynomial Approximate Likelihood (PAL), and bGPFA. The results demonstrate that ccGPFA is both more accurate and substantially faster than existing methods.

  • Superior Accuracy: The ccGPFA model (using a Negative Binomial likelihood) achieved a test Negative Log-Likelihood (NLL) of 0.3333, outperforming the best baseline, PAL (0.3407).
  • Massive Speedup: ccGPFA completed inference in just 7.67 seconds. In contrast, PAL took approximately 63 seconds, and the more complex bGPFA required over 1300 seconds.
  • Effective Latent Discovery: Qualitative analysis showed that the model successfully identified cyclic patterns corresponding to visual stimuli. Using an Automatic Relevance Determination (ARD) prior, the model automatically pruned 4 out of 10 latent dimensions, resulting in a concise and interpretable representation.

Discussion and Future Work

While ccGPFA offers a significant leap in performance, it shares common challenges with other factor analysis models, such as the inherent non-identifiability between loading weights and latent processes. The authors addressed this using standard orthonormalization post-processing. Future work may extend this conjugate framework to model interactions between multiple brain regions or utilize non-Euclidean manifolds to represent complex neural states. By bypassing high-variance Monte Carlo estimates, ccGPFA provides a robust and scalable tool for high-throughput neural data analysis.

Summary updated

BibTeX
@inproceedings{nadew2024conditionally,
  author = {Yididiya Y. Nadew and Xuhui Fan and Christopher J. Quinn},
  title = {{Conditionally-Conjugate Gaussian Process Factor Analysis for Spike Count Data via Data Augmentation}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=zgiT3uxvCF},
}
Figure