An Efficient and Flexible Spike Train Model via Empirical Bayes
This work addresses the challenge of accurately modeling neural spike data for neuroscientists, though it appears incremental as it builds on existing NB-GLM and empirical Bayes approaches.
The paper tackles the problem of overfitting and inaccurate parameter estimation in modeling neural spike responses, particularly for over-dispersed spiking behavior, by proposing a hierarchical parametric empirical Bayes method that integrates GLMs and empirical Bayes theory. The result shows that the method outperforms NB-GLM and Poisson-GLM in predictive log-likelihood on held-out data for retinal neurons.
Accurate statistical models of neural spike responses can characterize the information carried by neural populations. But the limited samples of spike counts during recording usually result in model overfitting. Besides, current models assume spike counts to be Poisson-distributed, which ignores the fact that many neurons demonstrate over-dispersed spiking behaviour. Although the Negative Binomial Generalized Linear Model (NB-GLM) provides a powerful tool for modeling over-dispersed spike counts, the maximum likelihood-based standard NB-GLM leads to highly variable and inaccurate parameter estimates. Thus, we propose a hierarchical parametric empirical Bayes method to estimate the neural spike responses among neuronal population. Our method integrates both Generalized Linear Models (GLMs) and empirical Bayes theory, which aims to (1) improve the accuracy and reliability of parameter estimation, compared to the maximum likelihood-based method for NB-GLM and Poisson-GLM; (2) effectively capture the over-dispersion nature of spike counts from both simulated data and experimental data; and (3) provide insight into both neural interactions and spiking behaviours of the neuronal populations. We apply our approach to study both simulated data and experimental neural data. The estimation of simulation data indicates that the new framework can accurately predict mean spike counts simulated from different models and recover the connectivity weights among neural populations. The estimation based on retinal neurons demonstrate the proposed method outperforms both NB-GLM and Poisson-GLM in terms of the predictive log-likelihood of held-out data. Codes are available in https://doi.org/10.5281/zenodo.4704423