Feature Collapse

Uncertainty estimation is an emerging field in trustworthy artificial intelligence for industrial cyber-physical systems (CPSs), as it ensures reliable detection of unprecedented situations, also known as out-of-distribution (OOD) samples. In this paper, we introduce Feature Collapse Mitigation (FCM), a lightweight method that utilizes a self-supervised reconstruction loss and Radial Basis Function neurons to provide predictive uncertainty. Unlike conventional approaches such as Monte Carlo (MC) Dropout, which require multiple forward passes, significantly increasing the computational cost, FCM offers efficient single forward-pass uncertainty estimation. The proposed approach reduces the intensity of the phenomenon of feature collapse in a self-supervised manner, which commonly degrades uncertainty estimates, without relying on prior knowledge of the in-domain (ID) data or OOD samples. Experiments on synthetic, benchmark and real world datasets demonstrate the potential and reliability of FCM in delivering reliable uncertainty estimation in various scenarios.