Unsupervised Detection of Cell Assemblies with Graph Neural Networks

• PAPER

Abstract

Cell assemblies, putative units of neural computation, manifest themselves as repeating and temporally coordinated activity of neurons. However, understanding of their role in brain function is hampered by a lack of scalable methods for their unsupervised detection. We propose using a graph neural network for embedding spike data into a sequence of fixed size vectors and clustering them based on their self-similarity across time. We validate our method on synthetic data and real neural recordings.

Quickstart

Install dependencies

We recommend starting with a fresh Conda environment and installing all the dependencies in requirements.txt (although not all of them are needed for the code in this repo to work). Assuming that your CUDA version is 12.1:

conda create -n graphseq python=3.9
conda activate graphseq
pip install -r requirements.txt

Generate a synthetic dataset

cd demo
python make_dataset.py \
    --p_drop 0.2 \
    --gap_ts 800 \
    --seqlen 100 \
    --jitter_std 15 \
    --winsize 100 \
    --step_sz 4 \
    --tau 25

Run sequence detection

cd demo
python detect.py \
    --snapshot_interval 5 \
    --z_dim 6 \
    --K 6 \
    --epochs 400

Visualize the optimization process

cd scripts
sh make_vids2.sh

You video will be assembled from frames (stored in data) and saved as an mp4 and a gif file.

Tensorboard

For your dataset, you might want to monitor the losses to tweak the hyperparameters. After at least one optimization run is started

cd runs
tensorboard --logdir=.

and navigate to http://localhost:6006 to monitor the losses.