Dis-inhibitory neuronal circuits can control the sign of synaptic plasticity

This repository contains the code associated with the NeurIPS 2023 paper "Dis-inhibitory neuronal circuits can control the sign of synaptic plasticity".

Installation

The following instructions should allow you to replicate our results in a fresh virtual environment. We used Python version 3.8.10.

1. Installing JAX with GPU support

Installing JAX with GPU support is required to replicate our results. You can use the following command:

pip install --upgrade "jax[cuda11_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

Note that the exact version you need might depend on the CUDA version on your local hardware. (see https://github.com/google/jax#installation for detailed instructions)

2. Other packages

All other required packages can be installed by running pip install -r requirements.txt

Replicating Results

1. Replicating figures

To replicate the experimental results we report in Figures 3 & 4 and Supplementary Figures S1 & S2, refer to the respective jupyter notebooks inside the notebooks folder.

2. Training multi-layer networks on computer vision benchmarks

For training of multi-layer networks on computer vision benchmarks, first go to the conf/dataset directory and change the path variable in fmnist.yaml and mnist.yaml to a local directory of your choice.

We use hydra for experiment logging and configuration. Simulations are run by calling the main.py file. The following commands train some of the networks we report in the manuscript:

• To train a 3-Layer MLP with classic backprop on Fashion-MNIST:

python main.py dataset=fmnist model/vf=non-dale trainer=bp model.vf.nb_hidden=3

• To train a 3-Layer MLP with dis-inhibitory control using the "exact inverse" learning rule:

python main.py dataset=fmnist model/vf=dale trainer=exact-inv model.vf.nb_hidden=3

• To train a 3-Layer MLP with dis-inhibitory control using the "linear threshold" learning rule:

python main.py dataset=fmnist model/vf=dale trainer=lin-thresh model.vf.nb_hidden=3

To train networks on MNIST instead of Fashion-MNIST, simply change fmnist to mnist in the commands above. To use the average Jacobian as feedback weights, add trainer.average_fb_weights=True to the commands.

All other simulation parameters can be be overridden using the same syntax as above or changed in the conf directory (refer to the https://hydra.cc/ documentation for details on the configuration files).

Citation

@article{Rossbroich2023-DisinhControl,
      title={Dis-inhibitory neuronal circuits can control the sign of synaptic plasticity}, 
      author={Julian Rossbroich and Friedemann Zenke},
      year={2023},
      eprint={2310.19614},
      archivePrefix={arXiv},
      journal={arXiv}
      primaryClass={q-bio.NC}
}