Code for Phase of firing does not reflect temporal order in sequence memory of humans and recurrent neural networks

Link to preprint

Example usage

Install the conda environment and train an RNN model:

cd sequence-memory
conda env create -f sequence.yml
activate sequence
python rnn_model/rnn/run_single_model.py

Expected behavior: A trained RNN model performing a working-memory task should be obtained and saved.

Expected run-time: One single RNN model took on average around 5-6 hours to train on a Nvidea 2080-ti GPU.

Reproducing the paper Figures

Pull the model and data files from this repo, by first installing git lfs. Alternatively, retrain new RNNs using

python rnn_model/rnn/run_single_model.py

or

wandb sweep rnn_model/rnn/sweep.yml
rnn_model/rnn/run_sweep.py # after adding sweep ID to top of file

and obtain summary statistics over multiple models using

rnn_model/rnn/run_summary.py

Either-way, the figures can then be recreated by running the notebooks in: rnn_model/generate_figures.

Note

Code tested on a Ubuntu system with package versions given in the sequence.yml file