Energy ring improves recognition of known and unknown classes
Energy Ring: Stabilizing Feature Energy for Reliable Test-Time Discovery
Artificial Intelligence
Summary
Systems that recognize objects while running often struggle to spot new, unfamiliar items without confusing them for known ones. The authors found that previous methods mix features of known and unknown classes, making it hard to tell them apart during operation. They propose a new method called Energy Ring that keeps the known and unknown classes separated by stabilizing their energy levels, making detection more reliable. Their approach uses special techniques to keep known categories distinct and groups unknown items tightly, improving the system’s ability to adapt and predict correctly.
What this means in practice
- •For machine learning engineers: Enhance deployed recognition systems to better detect new unseen classes without retraining on new labels.
- •For security analysts: Improve anomaly detection in streaming data by reliably distinguishing unfamiliar patterns from known ones during real-time monitoring.
Authors
Tianle Liu, Fan Lyu, Tingliang Feng, Chenggong Ni, Tianqi Wang, Zhang Zhang, Fuyuan Hu, Rui Yao, Tao Zhou
Abstract
Test-Time Discovery (TTD) aims to recognize known classes and detect novel ones while the model is already deployed, operating directly on a continual unlabeled test-time stream. Existing TTD methods are overly conservative, making known and unknown features collapse together and difficult to separate during test-time adaptation. In this paper, we revisit TTD from an energy perspective and find that the natural energy gap between known and unknown classes is unstable and easily collapses. We therefore propose Energy Ring to stabilize this gap by anchoring the two types of classes on separate energy rings. A lightweight adapter enlarges the norm-based gap while preserving known-class directions, and a perturbation consistency module forms compact clusters for unknown classes. During testing, energy-guided confidence weighting improves prediction, and a stability-based rectification mechanism corrects unreliable updates. Experimental results show that Energy Ring significantly improves TTD performance against existing TTD methods.