Recurrent neural network models of multi-area computation underlying decision-making

This Python package trains and analyzes multi-area RNNs. It trains RNNs by expanding on the pycog repository (https://github.com/frsong/pycog). Our package was written using Python 2.7.17.

• Installation Guide
• Examples

Installation Guide:

This should take only a couple minutes on a local computer.

1. Create a virtual environment. We recommend using virtualenvrapper (https://virtualenvwrapper.readthedocs.io/en/latest/)

$ pip install virtualenvwrapper
$ source /usr/local/bin/virtualenvwrapper.sh
$ mkvirtualenv -p python2.7 your-virtual-env-name

You may have to change the path to virtualenvwrapper.sh depending where it was installed by pip.

2. Add multi-area-cleaned to path

$ add2virtualenv /path/to/multi-area-cleaned
$ add2virtualenv /path/to/multi-area-cleaned/pycog

3. Install the dependencies

In the main directory, run:

pip install -r requirements.txt

This step installs everything from jupyter notebook to matplotlib for plotting. Note these will be done inside your virtual environment and will not disrupt your base python installation.

Examples

To generate the figures for the paper, run the following Jupyter notebooks. For Revision_main.ipynb and Revision_exemplar.ipynb, most of the cells rely on saved data and can be run immediately. Generating the PSTHs and PC plots takes a bit of time. Dynamics_polished.ipynb should run in about 5 minutes. You may get occassional matplotlib warnings, but these can generally be ignored.

sims/Revision_main.ipynb
sims/Revision_exemplar.ipynb
sims/dynamics_polished.ipynb

These sims/Revision_main.ipynb uses saved values for the hyperparameter sweeps. The mutual information and dpca variance values are generated using the following files. To run these scripts, update the paths (path and rnn_datapath) in cfg_mk.py. Note that these scripts take a few hours to run.

./sims/get_mutualinfo_vals.sh
./sims/get_dpca_vals.sh

The null/potent values are generated using python sims/null_potent_dpca.py.

A network is trained by running:

python examples/do.py examples/models/2020-04-10_cb_simple_3areas.py train

The different RNN modelfiles are in examples/models/ and the trained models are in saved_rnns_server_apr/data/

Please send any questions about the code to michael.kleinman@ucla.edu