Information Bottleneck-Based Hebbian Learning Rule Naturally Ties Working Memory and Synaptic Updates

A biologically plausible implementation of the HSIC objective. Please cite our work if you use this codebase for research purposes:

@article{10.3389/fncom.2024.1240348,
	author = {Daruwalla, Kyle and Lipasti, Mikko},
	doi = {10.3389/fncom.2024.1240348},
	issn = {1662-5188},
	journal = {Frontiers in Computational Neuroscience},
	title = {Information bottleneck-based Hebbian learning rule naturally ties working memory and synaptic updates},
	url = {https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2024.1240348},
	volume = {18},
	year = {2024},
	bdsk-url-1 = {https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2024.1240348},
	bdsk-url-2 = {https://doi.org/10.3389/fncom.2024.1240348}
}

Setup instructions

To install the codebase, install Micromamba (or Conda), then run:

micromamba create -f environment.yaml
poetry install --no-root

Running the experiments

The code for running the experiments is located in the root directory with auxiliary code located in projectlib. Configuration files use Hydra and are found in configs.

Small-scale experiments

To run the reservoir experiment, activate the project environment, and run:

python train-reservoir.py

The results will be stored under outputs/train-reservoir.

To run the linear synthetic dataset experiment, activate the project environment, and run:

python train-lif-biohsic.py data=linear

The results will be stored under output/train-lif-biohsic.

To run the XOR synthetic dataset experiment, activate the project environment, and run:

python train-lif-biohsic.py data=xor

The results will be stored under output/train-lif-biohsic.

Large-scale experiments

To run the MNIST experiment, activate the project environment, and run:

python train-biohsic.py data=mnist

The results will be stored under output/train-biohsic.

To run the CIFAR-10 experiment, activate the project environment, and run:

python train-biohsic.py data=cifar10 model=cnn

The results will be stored under output/train-biohsic.

The back-propagation baselines can be run using the same commands but with the train-bp.py script.

Plot the results