Abstract
Out-of-distribution (OOD) detection is essential to deploying machine learning systems in the real world. However, the reliability of the existing OOD detectors is severely hampered when used in an environment with adversarial/natural perturbations. Being such a critical component, this necessitates the study of techniques to robustify it. In this work, we propose using the representa- tion learning power of self-supervision methods with better OOD scoring mechanism based on energy to improve the robustness of OOD detectors. Specifically, we propose a blend of flexible loss function formulations that can effectively learn robust features. Our findings merit the use of a new methodological perspective that focuses on robustifying OOD detection.
Description
The exciting success of deep machine learning models has made them the de facto choice as the solution to building intelligent systems. The recent progress in their performance has led to an increasing number of real-world applications being powered using deep learning. Some of these areas include autonomous cars, automated facial recognition for security, voice-controlled devices, etc. Many of these applications are critical to influence the lives of people and it is very important that the deployed models are reliable. One important aspect of enforcing reliability (Amodei et al., 2016) is to be able to detect out of distribution (OOD) data and prevent exposing the deep learning models to these. This makes OOD detection very crucial to deploying trustworthy machine learning models.
In this work, we propose to improve the robustness of OOD detection by augmenting it with popular self-supervision frameworks and using more meaningful OOD scoring function. We specifically explore the robustness property of contrastive losses like SimCLR and other self-supervision tasks like predicting geometric rotations. The idea is to extend the OOD detection learning with self-supervised component and further improve it using energy based scores. Some existing work has proposed solutions for adversarial robustness of self-supervision against perturbations. In the current work, we explore the transferability of robustness to a completely unknown dataset (OOD setting) under a variety of attacks ranging from natural OOD, natural corruptions to compositional attacks which are harder to detect.
Member
Sean Chung
sean.chung@wisc.edu Aditya Kumar Akash
aka@cs.wisc.edu Wissam Kontar
kontar@wisc.edu Shri Shruthi Shridhar
shridhar2@wisc.edu
