AutoMIND

Automated Model Inference from Neural Dynamics (AutoMIND) is an inverse modeling framework for investigating neural circuit mechanisms underlying population dynamics.

AutoMIND helps with efficient discovery of many parameter configurations that are consistent with target observations of neural population dynamics. To do so, it combines a flexible, highly parameterized spiking neural network as the mechanistic model (simulated in brian2), with powerful deep generative models (Normalizing Flows in pytorch) in the framework of simulation-based inference (powered by sbi).

For a sneak peak of the workflow and what's possible, check out the overview demo and our preprint, Deep inverse modeling reveals dynamic-dependent invariances in neural circuit mechanisms.

This repository contains the package automind, demo notebooks, links to generated simulation datasets and trained deep generative models (figshare link), as well as code to reproduce figures and results from the manuscript.

---

Running the code

After cloning this repo, we recommend creating a conda environment using the included environment.yml file, which installs the necessary conda and pip dependencies, as well as the package automind itself in editable mode:

git clone https://github.com/mackelab/automind.git
cd automind
conda env create -f environment.yml
conda activate automind

The codebase will be updated over the next few weeks to enable successive capabilities:

• Inference: sampling from included trained DGMs conditioning on the same summary statistics of example or new target observations:
  • Demo-1 on a synthetic target using PSD features.
  • Demo-2 on a real recording using network burst features, starting from spike timestamps.
• Training: training new DGMs on a different set of summary statistics or simulations.
• Parallel simulations: running and saving many simulations to disk, e.g., on compute cluster.
• Analysis: Analyzing and visualizing discovered parameter configurations.
• ...and more!

---

Dataset

Model parameter configurations, simulations, and trained deep generative models

Model configurations and simulations used to train DGMs, target observations (including experimental data from organoid and mouse), hundreds of discovered model configurations and corresponding simulations consistent with those targets, and trained posterior density estimators can be found on figshare.

See here for details and download instructions.