Tailoring Mixup to Data for Calibration

Official code for Similarity Kernel Mixup on classification tasks (see classification folder), and regression tasks (see regression folder). Each folder contains a separated README for information about setting up and running experiments. Code to reproduce experiments on toy dataset is included in the toy_datasets folder.

Taking into account similarity in Mixup

Introducing similarity into the interpolation is more efficient and provides more diversity than explicitly selecting the points to mix.

Similarity Kernel

Batch-normalized and centered Gaussian kernel

• Amplitude $\tau_{max}$ governs the strength of the interpolation
• Standard deviation $\tau_{std}$ governs the extent of mixing
• Stronger interpolation between similar points and reduce interpolation otherwise

Avoiding Manifold Intrusion

Reference

If you find this work interesting, or if this repository has been helpful to you, please consider citing our work as follows:

@article{bouniot2023tailoring,
  title={Tailoring Mixup to Data for Calibration},
  author={Bouniot, Quentin and Mozharovskyi, Pavlo and d'Alch{\'e}-Buc, Florence},
  journal={arXiv preprint arXiv:2311.01434},
  year={2023}
}