Bridging Cognitive Maps

This repository contains the code for the experiments and figures used in Bridging Cognitive Maps: a Hierarchical Active Inference Model of Spatial Alternation Tasks and the Hippocampal-Prefrontal Circuit by Toon Van de Maele, Bart Dhoedt, Tim Verbelen, and Giovanni Pezzulo.

If you find the code useful, please refer to our work using:

@misc{vandemaele2023bridging,
      title={Bridging Cognitive Maps: a Hierarchical Active Inference Model of Spatial Alternation Tasks and the Hippocampal-Prefrontal Circuit}, 
      author={Toon Van de Maele and Bart Dhoedt and Tim Verbelen and Giovanni Pezzulo},
      year={2023},
      eprint={2308.11463},
      archivePrefix={arXiv},
      primaryClass={q-bio.NC}
}

Installation

The experiments require Python 3.11 (we ran using Python 3.11.6). The easiest way to set up the requirements is by creating a virtualenvironment and running:

pip install -r requirements.txt

and install the repo in editable mode by running:

python setup.py develop 

The figures of the papers assume that trained models, and dataset are in the data folder, located in the project root. They can be downloaded using this link.

Running the experiments

The cognitive maps can be trained using the scripts in experiments/model-learning:

• The sequence dataset can be generated using the generate_full_explore_dataset.py script.
• The navigation model can be trained using the train_navigation_*.py script, where the * can be either a cscg or a hmm, depending on which model you need. This requires the dataset to be in the data/ folder.
• The navigation models can be evaluated using the resepective evaluate_planning_and_inference_*.py. Which will measure the success rate, when the agent starts in each possible pose in the maze, and is tasked to go to each of the corridors. NOTE: these scripts point to a location of a model to evaluate.
• The task model can be trained using the train_task_cscg_loc.py script.

The paper experiments and figures are located in experiments/figures. Where dedicated notebooks exist for each of the experiments.

Acknowledgments

The code for training the clone structured cognitive graphs comes from CSCG. The active inference implementation relies on PyMDP.