Source-linked AI summary
Energy Ring: Stabilizing Feature Energy for Reliable Test-Time Discovery
Tianle Liu, Fan Lyu, Tingliang Feng, Chenggong Ni, Tianqi Wang, Zhang Zhang, Fuyuan Hu, Rui Yao, Tao Zhou
TL;DR
In TTD, models must recognize known classes and discover novel ones in a continual unlabeled stream, but the energy gap between them can be unstable. Energy Ring uses separate energy rings and test-time mechanisms to stabilize that gap; it improves agreement performance over DAA and outperforms across four datasets in adaptability and stability. The current energy cycle still relies on preset energy tracks.
Problem
In TTD, the energy patterns of known and unknown classes can overlap and drift as novel classes arrive, making them difficult to distinguish reliably.
Source: §3.1 · p. 4Method
Energy Ring anchors known and unknown classes on separate energy rings, using an adapter and perturbation consistency to enlarge the gap and stabilize unknown-class clusters.
Source: §1 · p. 2Results
Energy Ring improves agreement performance by 3.33% compared to DAA and outperforms in adaptability and stability across four datasets.
Source: §1 · p. 2, §5 · p. 10Takeaways & Limitations
Energy Ring enables instant recognition of unknown classes while safeguarding accuracy on previously learned ones.
Source: §5 · p. 10The current energy cycle still relies on pre-set energy tracks.
Source: §5 · p. 10
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Test-Time Discovery (TTD) must recognize known classes and discover novel ones in a continual unlabeled stream after deployment, but existing methods can limit adaptation or collapse feature variation. Energy Ring stabilizes the energy gap between known and unknown classes through separate energy rings and reports improved performance over prior TTD methods.
1 INTRODUCTION
- TTD handles a continual unlabeled stream in which known and unknown samples arrive sequentially, requiring real-time recognition and discovery.Unlike Novel Class Discovery, TTD does not assume a fixed unlabeled set is available in advance.Source: §1 · p. 1
- Figure 1 contrasts HM's fixed spatial energy and DAA's single hypersphere with Energy Ring's separate rings and retained separability.The paper says DAA's shared feature norm may crowd and confuse known and unknown classes.Source: §1 · p. 1, Figure 1 · p. 2
- The natural energy gap between known and unknown classes can erode as unknown-class energy becomes chaotic, making the system unreliable.The paper describes known classes as often having higher energy and unknown classes as often having lower energy.Source: §1 · pp. 1–2
- Energy Ring improves agreement performance by 3.33% compared to DAA.The paper also reports clear improvements over prior TTD methods.Source: §1 · p. 2
- Energy Ring anchors known and unknown classes on separate energy rings, enlarging their energy separation while preserving known-class feature directions.Its perturbation-consistency module encourages compact, stable clusters for unknown samples.Source: §1 · p. 2
2 RELATED WORK
Related work spans novel-class discovery, test-time adaptation, and methods that combine discovery with deployment-time learning. The paper distinguishes TTD by its continual-stream setting and summarizes approaches with different trade-offs in adaptation and feature preservation.
2 RELATED WORK
- Novel Category Discovery separates representation learning from clustering unlabeled data, while On-the-fly Category Discovery hashes incoming images but keeps its backbone frozen.OCD does not fine-tune on new classes.Source: §2 · p. 2
- Test-Time Training adapts a pretrained model to distribution shifts online, but most methods focus on known classes and leave emerging categories unaddressed.Representative approaches use self-supervised objectives to update the backbone from the test stream.Source: §2 · p. 2
- TTD combines retaining old-class accuracy with discovering and learning new classes during inference in a continuous unlabeled stream.The task setting begins with pretraining on known classes and allows both known and unknown classes at deployment.Source: §2 · p. 3, Figure 2 · p. 3
- The first TTD implementation uses frozen-network external memory, whereas DAA updates a small set of adapters in real time.The frozen approach avoids gradient updates but forfeits feature-space refinement when old and new classes overlap.Source: §2 · p. 3
3 METHODS
Energy Ring addresses unstable separation between known and unknown features by placing them on distinct energy rings, then using those rings to guide prediction and test-time rectification.
3.1 TTD DEFINITION AND ENERGY-BASED DISCOVERY
- Because norm and angular differences can overlap or drift as novel classes arrive, the method constructs separate energy rings to stabilize known–unknown separation.Known classes occupy the higher-energy ring and unknown classes the lower-energy ring.Source: §3.1 · p. 4
3.2 PRECONSTRUCT ENERGY RINGS THROUGH WARM-UP
- Warm-up mixes features with noise as unknown proxies, then enforces outer and inner norm rings, preserves known-feature directions, and clusters norm-preserving unknown perturbations.Perturbations of the same unknown sample are encouraged to stay close, while those from different samples are pushed apart.Source: §3.2 · pp. 4–5
3.3 ENERGY RING GUIDED PREDICTION AND DISCOVERY
- Energy position adjusts class similarities within a narrow range, and a confidence threshold determines whether a sample is assigned to an existing or new class.Features closer to the outer ring receive slightly higher similarity toward known classes; the energy prior is intended to nudge, not dominate, prediction.Source: §3.3 · p. 5
- For predicted unknowns, perturbation stability identifies low-stability samples for rectification and test-time training, helping reduce prototype drift.Samples selected for rectification are added to test-time training and can serve as the first instance of a newly seen class.Source: §3.3 · p. 6
3.4 TEST-TIME TRAINING TO STABILIZE ENERGY RINGS
- Test-time updates preserve known-class representations while adapting to stable unknowns, and replay reprocesses each discovered class’s least-stable samples.The testing losses include pseudo-label learning and memory-based replay for known and seen classes.Source: §3.4 · p. 6
4 EXPERIMENT
The experiments evaluate Energy Ring on four known/unknown-class benchmarks using real-time discovery, post-test clustering, and analyses of its energy-based components. Results favor the method's balance across discovery and retention, while component studies examine the effects of weighting, perturbations, and ring settings.
4.1 EXPERIMENTAL DETAILS
- Experiments use four datasets split 7:3 into known and unknown classes, with a pretrained ViT-B/16 and TA, CA, known-accuracy, forgetting, and clustering metrics.The datasets are CIFAR100D, CUB-200D, Tiny-ImageNetD, and AircraftD; TA and CA assess discovery, while KA and KF assess known-class performance.Source: §4.1 · pp. 6–7
4.2 MAJOR COMPARISONS
- Energy Ring achieves more consistent TA and CA performance across four datasets, balancing discovery with retention of previously learned knowledge.Table 1 compares four datasets under real-time and post evaluation; real-time scores accumulate across test batches, while post evaluation reassesses all test samples.Source: Table 1 · p. 7, §4.2 · p. 8
- Energy Ring also has better post-evaluation clustering performance than competing methods across the four datasets.TA and CA reflect online discovery quality and class consistency; clustering metrics provide compatibility with post-hoc NCD evaluation.Source: §4.2 · p. 8, Table 2 · p. 8
- Energy Ring is more balanced across TA and CA, with larger red areas and results closer to the TA = CA balance line.The paper notes that high values on only one of the two complementary metrics do not necessarily indicate good overall classification performance.Source: §4.2 · p. 8, Figure 5 · p. 8
4.3 ANALYSIS ON ENERGY RING
- Simulated unknown classes have lower feature norms and larger average confidence differences, while excessive energy weighting lets norm dominate and collapses accuracy.Only a small modulation factor is required because feature-to-category similarities overlap around 0.7.Source: §4.3 · p. 8, Table 3 · p. 9
- Energy-weighted prediction improves the fixed model, and adding perturbation rectification yields TA+CA 81.18 real-time and 76.66 post; too few or excessive perturbations are ineffective or costly.The perturbation analysis reports that insufficient perturbations make stability judgments unreliable, while excessive perturbations add time cost without significant improvement.Source: §4.3 · p. 9, Table 4 · p. 9, Figure 6 · p. 9
- During warm-up, the ring radii are coupled; too small an unknown-class radius pushes the inner ring near the origin and destabilizes clusters, so known-class feature norm is fixed at 70.The fixed value is based on backbone known-class feature norms centering around 70.Source: §4.3 · p. 9, Figure 7 · p. 9
- Angular-space T-SNE on CIFAR100D (5 known and 5 unknown classes) shows clearer boundaries; CIFAR-100D (90+10) matching shows more novel classes found and fewer known samples assigned unknown.The visual comparison also describes chaotic predictions from base-model updating and missed categories with DAA.Source: §4.3 · p. 9, Figure 8 · p. 10, Figure 9 · p. 10
5 CONCLUSION
Energy Ring is presented as a TTD framework for recognizing unknown classes while safeguarding known-class accuracy. Experiments on four datasets show gains in adaptability and stability, though the current method relies on preset energy tracks.
5 CONCLUSION
- Energy Ring aims to recognize unknown classes while safeguarding accuracy on previously learned classes, and outperforms in adaptability and stability on four datasets.The conclusion describes its warm-up and test-time computation overhead as marginal compared with discovery-performance gains.Source: §5 · p. 10
- The lightweight adapter places features on concentric energy rings, preserves known-class directions, enlarges the energy gap, and forms compact, stable unknown-class clusters.Source: §5 · p. 10
- The current energy cycle relies on preset energy tracks; future work will explore dynamically setting them and extending to multimodal and continual learning.Source: §5 · p. 10