Feature Extraction for Out of Distribution Detection via Self-Supervised Learning
Claire Thorp, Sean Sisti, Walter Dean Bennette · 2022
In an operational environment, an automated system performing image classification can encounter observations belonging to classes for which it was not previously trained to recognize. In such cases, modern machine learning classifiers are prone to misclassifying Out Of Distribution (OOD) observations as one of the model’s known classes - with high confidence. In defense applications, such an error could have catastrophic consequences, and requires reliable OOD detection to mitigate this risk. Techniques exist to reliably detect OOD observations, but these techniques largely assume the availability of a useful OOD exposure set during model training. Unfortunately, such an exposure set cannot be assumed for many defense applications that suffer from low amounts of labeled (and unlabeled) data. Therefore, in this work we investigate enabling OOD detection without an OOD exposure set. Specifically, we investigate Self-Supervised Learning as a feature extraction method, with the hypothesis that Self-Supervised Learning can provide a rich feature set for OOD detection, without OOD exposure during training.