Skip to content

Official code repository for the publication "Latent Equilibrium: A unified learning theory for arbitrarily fast computation with arbitrarily slow neurons"

License

Notifications You must be signed in to change notification settings

unibe-cns/le_NeurIPS_code

Repository files navigation

Latent Equilibrium: A unified learning theory for arbitrarily fast computation with arbitrarily slow neurons

This repository contains the code to reproduce the results of the NeurIPS 2021 submission "Latent Equilibrium: A unified learning theory for arbitrarily fast computation with arbitrarily slow neurons" (also available on arXiv).

Requirements

To install requirements:

pip install -r requirements.txt

Training & Evaluation

Code for FC MNIST experiments (Fig.2b and 4ac)

The code can be found in fig2b_fig4ac_mnist/src/.

Running the experiments: For example, in order to run all the experiments needed to reproduce Fig. 2b, execute:

cd fig2b_fig4ac_mnist/src/
/bin/bash 2b_jobs.sh

The results of each run, that is for example metrics, output and configurations, will be saved in fig2b_fig4ac_mnist/runs/{run_number}/.

For the experiment in Fig.4 replace 2b_jobs.sh with 4a_jobs.sh or 4c_jobs.sh respectively

The seeds chosen for these experiments were 42 69 12345 98765 38274 28374 42848 48393 83475 57381.

Code for HIGGS, MNIST and CIFAR10 with and without LE (Fig. 2cde).

The code can be found in fig2cde_higgs_mnist_cifar10.

The code configuration is integrated into the main files and only a few parameters are configured via argparse.

To run the code, check the respective submit_python_*_v100.sh file which contains examples and all run configurations for all seeds used.

The seeds chosen for these experiments were 1, 2, 3, 5, 7, 8, 13, 21, 34. (Fibonacci + lucky number 7), resulting in 9 seeds for each experiment.

Results can be found in the respective log file produced from the std out of the running code via python -u *_training.py > file.log.

Code for Dendritic Microcircuits with and without LE (Fig.3 and 5)

The code can be found in fig3fig5_dendritic_microcircuits/src/.

The experiments are configured using config files. All config files required for the production of the plotted results are in fig3fig5_dendritic_microcircuits/experiment_configs/. The naming scheme of the config files is as follows {task name}_{with LE or not}_tpres_{tpres in unit dt}.yaml where task name is bars (Fig.3) or mimic (Fig.5) and with LE or not is either le or orig.

For each run the results will be saved in fig3fig5_dendritic_microcircuits/experiment_results/{config file name}_{timestamp}/.

To run an experiment:

cd fig3fig5_dendritic_microcircuits/src/
python3 run_bars.py train ../experiment_configs/{chosen_config_file}

For the experiment in Fig.5 replace run_bars.py with run_single_mc.py

To plot the results of a run:

cd fig3fig5_dendritic_microcircuits/src/
python3 run_bars.py eval ../experiment_results/{results_dir_of_run_to_be_evaluated}

This will generate plots of the results (depending on how many variables you configured to be recorded, more or less plots can be generated) and save them in the respective results directory. Which plots are plotted is defined in run_X.py

Reproduce all data needed for Fig3:

For the results shown in Fig.3 all config files with the name bars_*.yaml need to be run for 10 different seeds (configurable in the config file). The seeds chosen for these experiments were 12345, 12346, 12347, 12348, 12349, 12350, 12351, 12352, 12353, 12354.

Contributing

📋 TODO: Pick a licence and describe how to contribute to your code repository.

About

Official code repository for the publication "Latent Equilibrium: A unified learning theory for arbitrarily fast computation with arbitrarily slow neurons"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published