A simple Training-Free Method for Rejection Option

Eduardo Dadalto Câmara Gomes, Marco Romanelli, Federica Granese, Pablo Piantanida

Keywords: rejection option, safety AI, deep learning

TL;DR: We present a simple yet effective method to implement the rejection option for a pre-trained classifier.

Abstract: We present a simple yet effective method to implement the rejection option for a pre-trained classifier. Our method is based on a sound mathematical framework, enjoys good properties, and is hyperparameter free. It is lightweight, since it does not require any re-training of the network, and it is flexible, since it can be used with any model that outputs soft-probabilities. We compare our solution to state-of-the-art methods considering popular benchmarks (Cifar-10, Cifar-100, SVHN), and various models (VGG-16, DenseNet-121, ResNet-34). At evaluation time, our method, which is applied post-training to any classification model, achieves similar or better results with respect to its competitors that usually require further training and/or tuning of the models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)