Exploring the dynamical repertoire of GLM network, properties of stochastic spiking networks, and tasks for RNNs
Function and scripts to generate recurrent neural network (RNN) constituted by generalized linear models (GLM); then perform inference on parameters given target spikng patterns or readout time series. Some main scripts I am working on:
-
GLM_RNN.py- Is a simple demonstration for generative model with spikes either from ground-truth parameters or latent time seires. We then perform inference and seek to reconstruct similar spiking patterns. -
brunel_GLM.py- Another demonstration for different target spiking patterns (AI, AR, SI, SR) described in the Brunel 2000 paper. We then perform inference with GLM network to capture these population spiking patterns. -
lowrank_GLM.py- Simulates spikes with low-rank connectivity and compares inference performance with or without low-rank regularization. -
RSNN.py.py- Implementing Bellec et al. 2022 method of pseudo-gradient and back-prop through time to train GLM networks with perceptual decision making tasks. -
GFORCE_seq.py- Demo for GLM-FORCE algorithm with target that is sequential spiking pattern.
Currently working on packing functions/classes for this:
from GLM_RNN.GLMnet import nonlinearity
nonlinearity()
Scripts for FORCE learning algorithm and its generalization to GLM networks, different target dynamics, and network constraints.
Preliminary tests for full-FORCE learning algorithm and its generalization to GLM networks. The goal is to jointly learn latent 'hints' for training.
Still working on examples for unit-test and packaging.