Classification Reconstruction Encoder

Overview

This repository hosts the official implementation of the paper "Classification and Reconstruction Processes in Deep Predictive Coding Networks: Antagonists or Allies?" Within this repository, we provide the implementation of the primary tool discussed in the paper, the Classification Reconstruction Encoder (CRE), along with the necessary training scripts.

Requirements

• Python 3.9.6.
• timm 0.9.5
• torch 2.0.1
• torchvision 0.15.2
• numpy 1.25.1

Installation

Clone the repository to your local machine and install the requirements.

Usage

To train a FC-based or CNN-based CRE, run the training script train_cre.py. To train a ViT-based CRE, run the training script train_cre_vit.py. Model and training parameters can be changed in both training scripts.

Citation

@Article{rathjens2024classification,
      title={Classification and Reconstruction Processes in Deep Predictive Coding Networks: Antagonists or Allies?}, 
      author={Jan Rathjens and Laurenz Wiskott},
      year={2024},
      eprint={2401.09237},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Contact

For questions and feedback, contact us at [email protected].

License

This project is under the BSD-3-Clause license. See LICENSE for details.