Cerebellar-driven cortical networks. A cortical RNN projects to a cerebellar network, which in turn returns feedback to the RNN (https://www.biorxiv.org/content/10.1101/2022.11.14.516257v1).
Beyond the standard python libraries, you will need Pytorch (we use version 1.7.0, but later should work) to define the neural network models and as well as ignite (version 0.4.2) which wraps the training regime. Sacred (version 0.7.4) is used to record experiment details.
Scripts to run the tasks presented in the paper can be found in the /scripts folder. For a given script:
- Define model/training/dataset hyperparameters*. These can be changed in the respective /configs folder or in the script itself (as a config update).
- Define the OBSERVE variable (preset to False). If true, then experiment results will be saved in a generated /sims folder.
- Run the script
*For the consolidation script (in /delass) it's only necessary to define the path to the pretrained model. At the moment this points to /delass/sims/1.