Discrete Dictionary-based Decomposition Layer for Structured Representation Learning

Source code for the NeurIPS'24 paper: Discrete Dictionary-based Decomposition Layer for Structured Representation Learning

TL;DR: Discrete dictionary-based decomposition that maps given data to the structured representations of tensor product representation (TPR) framework, thereby improving systematic generalization of TPR-based models.

Requirement

• Python 3.8
• PyTorch 1.13

pip install -r requirements.txt

Acknowledement

We've developed our code using open-source implementations for various tasks, as below.

• AID

• bAbI task (FWM)

• bAbI task (TPR-RNN)

• Sort-of-CLEVR task

• WikiText-103 task

Citation

@article{park2024discrete,
  title={Discrete Dictionary-based Decomposition Layer for Structured Representation Learning},
  author={Park, Taewon and Kim, Hyun-Chul and Lee, Minho},
  journal={arXiv preprint arXiv:2406.06976},
  year={2024}
}