colinbredenberg/oneirogen-hypothesis

Code repository for the paper "The oneirogen hypothesis: modeling the hallucinatory effects of classical psychedelics in terms of replay-dependent plasticity mechanisms"

Codebase for the paper: "The oneirogen hypothesis: modeling the hallucinatory effects of classical psychedelics in terms of replay-dependent plasticity mechanisms"

By Colin Bredenberg, Fabrice Normandin, Blake Richards, and Guillaume Lajoie

This code repository was produced by simplifying and adapting the Mila IDT team's BeyondBackprop library: https://github.com/mila-iqia/BeyondBackprop (still under active development), keeping only code required for reproducing this paper.

Also inspired / adapted from https://www.github.com/ernoult/ScalingDTP as well as https://github.com/ashleve/lightning-hydra-template

Installation

Step 1: Create a Python 3.9 virtual envioronment.

Step 2: Within that environment, in the command line, run:

pip install -e .

Codebase structure

The main logic of our method is in beyond_backprop/algorithms/algorithm.py

This Algorithm class is based on the LightningModule class of PyTorch-Lightning.

Here is how the codebase is roughly structured:

├── main.py                # training script
└── beyond_backprop
    ├── algorithms     # Learning algorithm implementations
    ├── configs        # configuration files and classes for Hydra
    ├── datamodules    # Datasets and dataloaders
    ├── networks       # Network definitions
    └── utils          # Utility functions

Running the code

This codebase uses Hydra for configuration management and for the command-line interface. You can see all the command-line groups available with:

python main.py --help

From there, you can see all the available options for each group. The three most important groups are "algorithm", "datamodule", and "network".

You can select a configuration from each group, as well as overwrite any of the options in the configuration.

python main.py algorithm=backprop datamodule=mnist trainer.max_epochs=10

Running experiments

To rerun the experiments used to produce the main paper figures, choose an experiment in the beyond_backprop/configs/experiment folder and from the oneirogen-hypothesis folder run:

python main.py experiment=[experiment_name]

experiment_name options:

Multicompartment model trained on MNIST (main results): rm_wake_sleep_ff Multicompartment model trained on CIFAR10: rm_wake_sleep_ff_cifar No learning control: rm_wake_sleep_ff_no_learning Single compartment model: rm_wake_sleep_ff_no_batch_norm Noise-based hallucination protocol: rm_wake_sleep_ff_noise Recurrent network model: rm_wake_sleep (note: this model is much larger and takes longer to run)

Expected runtime: 2-3 hours. Generating full hallucination videos may take longer, but experiment results will appear in your beyond_backprop/logs/ folder before then.

If you do not have a gpu on your local machine, you may need to change the line:

- override /trainer: default

in the config file you would like to run to:

- override /trainer: cpu