On Path Integration of Grid Cells: Group Representation and Isotropic Scaling

This repo contains the official implementation for the paper On Path Integration of Grid Cells: Group Representation and Isotropic Scaling (NeurIPS 2021)

by Ruiqi Gao, Jianwen Xie, Xue-Xin Wei, Song-Chun Zhu, and Ying Nian Wu

Our model learns clear hexagon grid patterns of multiple scales which share observed properties of the grid cells in the rodent brain, by optimizing a simple loss function:

The learned model is also capable of accurate long distance path integration:

Dependencies

• Python >= 3.5

Run the following to install a set of python packages necessary for running the code:

pip install -r requirements.txt

Usage

Train and evaluate our model through main.py.

python3 main.py
  --mode: <train|visualize|path_integration|error_correction> 
    (running mode: train / visualize filters / path integration / error correction)
  --ckpt: ckeckpoint file to load
    (default: None)

• For training the model from scratch, set --mode=train and --ckpt=None.
• For the other three modes, the path of a ckeckpoint file is required to set to --ckpt.

References

If you find the code useful for your research, please consider citing

@article{gao2020path,
  title={On Path Integration of Grid Cells: Group Representation and Isotropic Scaling},
  author={Gao, Ruiqi and Xie, Jianwen and Wei, Xue-Xin and Zhu, Song-Chun and Wu, Ying Nian},
  journal={arXiv preprint arXiv:2006.10259},
  year={2020}
}

This work is built upon a previous paper which might also interest you:

• Gao, Ruiqi, Jianwen Xie, Song-Chun Zhu, and Ying Nian Wu. "Learning grid cells as vector representation of self-position coupled with matrix representation of self-motion." International Conference on Learning Representations, 2019.