DeepVoro: Few-shot Learning as Cluster-induced Voronoi Diagrams (ICLR 2022)

Few-shot Learning as Cluster-induced Voronoi Diagrams: A Geometric Approach

The Tenth International Conference on Learning Representations (ICLR 2022)

OpenReview

arXiv

Introduction

In metric-based Few-Shot Learning (FSL), the classification model (e.g. Prototypical Networks and Logistic Regression) can be intepreted as either a Voronoi diagram (VD), or a Power diagram (PD) constructed in the feature space.

In this work we dive deeper, geometrically, into this line and extend VD/PD to Cluster-induced Voronoi Diagrams (CIVD), proposed in FOCS 2013 and SIAM Journal on Computing 2017 (PDF), and establish DeepVoro to combat the extreme data insufficiency in FSL.

Prerequisites

Install dependencies via:

pip install -r requirements.txt

Then please follow S2M2_fewshot to prepare the datasets and download the pretrained models, after which your directories should be like:

checkpoints/miniImagenet/WideResNet28_10_S2M2_R/470.tar checkpoints/CUB/WideResNet28_10_S2M2_R/255.tar checkpoints/tieredImageNet/WideResNet28_10_S2M2_R/199.tar

Reproducing the results

In this tutorial we use CUB dataset, the smallest among three, to illustrate the DeepVoro workflow (there should have been a directory "filelists/CUB/CUB_200_2011" storing the CUB images, after the previous step).

Step 1: Feature extraction:

bash run_save_plk_cub.sh

Step 2: Establish individual VDs:

bash run_individual_vds.sh

Step 3: Establish CIVD: see jupyter notebook "deepvoro.ipynb" for details; you should be able to reproduce the following results:

References

If you find the software useful please consider citing:

@misc{ma2022fewshot,
      title={Few-shot Learning as Cluster-induced Voronoi Diagrams: A Geometric Approach},
      author={Chunwei Ma and Ziyun Huang and Mingchen Gao and Jinhui Xu},
      year={2022},
      eprint={2202.02471},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

or

@inproceedings{ma2022fewshot,
    title={Few-shot Learning via Dirichlet Tessellation Ensemble},
    author={Chunwei Ma and Ziyun Huang and Mingchen Gao and Jinhui Xu},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=6kCiVaoQdx9}
}

Acknowledgments

In this project we use (parts of) the official implementations of the following works:

• S2M2_fewshot
• simple_shot
• Few_Shot_Distribution_Calibration

We thank the respective authors for open-sourcing their methods, which makes this work possible.

If you have any problem please contact me.