Sampling-based inference in a population of Poisson spiking neurons

• Overview
• User guide and system requirements
• License

Overview

This code simulates a recurrent neural network and coupled networks of Poisson spiking neurons which implement sampling-based Bayesian inference to approximate the multivariate posterior of latent stimuli. In the network model, the sampling is driven by the internal Poisson variability of spike generation, and the structured recurrent connections store a prior of latent stimuli. We show the fixed (discrete) Poisson variability is able to sample different types of continuous distributions with different uncertainties.

The network has excitatory (E) and inhibitory (I) neurons, with each neuron modeled as a Hawkes process. The E neurons have Gaussian tunings over the stimulus feature, while the I neurons don't have selectivity over stimulus and only keep the stability of the network.

Recurrent connections in the network

• Two E neurons are connected according to their tuning similarity.
• All other connections in the network, including E to I, I to E and I to I are unstructured.

More details of the network model and the results of this code can be found at our bioRxiv paper at: https://www.biorxiv.org/content/10.1101/2020.11.18.389197v2

User guide and system Requirements

The code was developed on MATLAB R2018b and can be ran on MATLAB whose version is at least R2018b, no matter what operating system. Directly download the whole package of code.

There are four main codes the user can use to reproduce the results in this study

• simExpMdls.m
• Cluster_code\simNet_ScanPars.m
• Cluster_code\ansNetDat_ScanPars.m
• Cluster_code\scanCoupledNet_RandPars_Cluster.m

Guideline of running simExpMdls.m

Directly run simExpMdls.m and you can get most of demo results in the presented study.

You can change the variable flagTask in simExpMdls.m to get the results under different demo tasks.

The simExpMdls.m can be ran on a standand computer with enough RAM. The computer should have at least 8GB of RAM.

Guideline of running Cluster_code\simNet_ScanPars.m

The simNet_ScanPars.m should be ran on a high performance cluster (HPC).

A single loop in simNet_ScanPars.m will take about 4GB of RAM and ran about 40 mins in a single thread.

The user may need to revise simNet_ScanPars.m based on the job scheduler or mangement on the HPC.

Guideline of running Cluster_code\ansNetDat_ScanPars.m

ansNetDat_ScanPars.m should be only ran after running Cluster_code\simNet_ScanPars.m to obtain a results.

You need to set fileName in ansNetDat_ScanPars.m to read the results ran by simNet_ScanPars.m.

License

This project is covered under the Apache 2.0 License.