Code corresponding to the NeurIPS 2020 paper Reverse-engineering Recurrent Neural Network solutions to a hierarchical inference task for mice.

After cloning the repository, run pip install -r requirements.txt to install the project's dependencies. We used Python 3 and did not test Python 2.

Our code was written to run on local machine or on a SLURM cluster. There are two main scripts to run.

1. train.py: This will train a RNN with parameters specified inside utils/params.py. During training, two types of objects will be written to disk inside a newly created runs/ directory. The first is a TensorBoard file which logs basic information (e.g. loss, average number of RNN steps per trial, average fraction of correct actions per trial). The second are model checkpoints through the training process.

2. analyze.py: This will take a trained RNN and generate all the plots contained in the paper. You'll need to specify the train run id (e.g. rnn, block_side_probs=0.8, max_stim_strength=2.5_2020-06-18 11:14:27.427969) inside analyze.py. The plots will be written to disk inside that run directory, in a newly created directory named analyze. A PDF containing all the images will also be generated, in case you want to send them all to someone simultaneously.

If you want to run on a SLURM cluster, use the ann-rnn.sh bash script.

• If you want to run analyze, depending on how many blocks you want to average over, the run may take a very long time. This is likely due to fitting the traditionally distilled model (see utils.analysis.distill_model_traditional()). However, the model will be written to disk so that subsequent runs will take less time.

• Some of the variable names are inconsistent throughout the repo because our understanding evolved as the project progressed. Sadly, I haven't had time to rename variables for reader clarity.

• There are also remnants of directions we never explored. For instance, we considered playing with different parameter initializations and connectivity constraints (masks) but these were never actually used.

We'd be delighted to hear from you. Email Rylan Schaeffer at rylanschaeffer@g.harvard.edu and cc Ila Fiete at fiete@mit.edu.