On Outlier Exposure with Generative Models

This is the code for the workshop paper On Outlier Exposure with Generative Models as published on the NeurIPS 2022 Machine Learning Safety Workshop.

In the paper, we take a first step into exploring the idea of using generative models to sample example outliers from low likelihood regions of the data distribution. Such samples can be used as synthetic outliers for outlier exposure, and we show that they tend to make the model more robust to Out-of-Distribution (OOD) data compared to models exposed to uniform noise.

Dependencies

• pytorch
• torchvision
• torchtext
• pytorch-ood

Citation

@inproceedings{kirchheim2022outlier,
  title={On Outlier Exposure with Generative Models},
  author={Kirchheim, Konstantin and Ortmeier, Frank},
  booktitle={NeurIPS ML Safety Workshop},
  year={2022}
}