Unsupervised Abstract Reasoning for Raven’s Problem Matrices

This code is the implementation of our TIP paper.

This is the first unsupervised abstract reasoning method on Raven’s Progressive Matrices, it is an extention of our arxiv preprint.

Comparision with some supervised methods.

Average testing accuracy on the RAVEN, I-RAVEN, and PGM dataset

Method Raven I-RAVEN PGM
CNN 36.97 13.26 33.00
ResNet50 86.26 – 42.00
DCNet (ICLR2021) 93.58 49.36 68.57
NCD (Ours) 36.99 48.22 47.62

Generalization test results on PGM dataset

Method neutral interpolation extrapolation
WReN (ICML2018) 62.6 64.4 17.2
DCNet (ICLR2021) 68.6 59.7 17.8
MXGNet (ICLR2020) 89.6 84.6 18.4
NCD (Ours) 47.6 47.0 24.9

Citation

If our code is useful for your research, please cite the following papers.

@article{zhuo2021unsup,
  title={Unsupervised Abstract Reasoning for Raven’s Problem Matrices},
  author={Tao Zhuo, Qiang Huang, and Mohan Kankanhalli},
  journal={IEEE Transactions on Image Processing},
  year={2021}
}


@article{zhuo2020solving,
  title={Solving Raven's Progressive Matrices with Neural Networks},
  author={Tao Zhuo and Mohan Kankanhalli},
  journal={arXiv preprint arXiv:2002.01646},
  year={2020}
}


@inproceedings{iclr2021,  
    author={Tao Zhuo and Mohan Kankanhalli},  
    title={Effective Abstract Reasoning with Dual-Contrast Network},  
    booktitle={International Conference on Learning Representations (ICLR)},      
    year={2021}
}

GitHub

https://github.com/visiontao/ncd