acceptodds
Under review as a conference paper at ICLR 2027

SPARK: Structural Priors Activated from closed-set Recognition Knowledge for On-the-Fly Category Discovery

Abstract

On-the-Fly Category Discovery aims to identify known classes and continuously distinguish new classes from a streaming test sequence. Existing methods often entangle category discovery with backbone optimization by introducing discovery-specific modules or objectives that reshape the closed-set feature space. In this paper, we ask a simpler question: can a standard closed-set classifier already provide useful structural priors for online discovery? We answer this question affirmatively and propose SPARK, a plug-and-play framework that activates the geometric knowledge encoded by closed-set supervision. Instead of treating known classes as isolated prototypes, SPARK learns Local Manifold Anchors around each known class. These anchors preserve inter-class discrimination while providing non-redundant local coverage of intra-class semantic variations. The union of all anchors forms a discrete approximation to a shared superclass manifold. At test time, when a streaming sample is rejected as unknown, SPARK analytically initializes its corresponding prototype under the constraint of this manifold and then updates the prototype online as more samples arrive. Experiments on seven public benchmarks show that SPARK achieves strong overall performance. These results suggest that on-the-fly category discovery does not necessarily require heavy specialized modules and the structural priors already present in closed-set classifiers can be directly activated for open-world recognition.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Category Discovery, On-the-Fly Category Discovery, Manifold Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LFs5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPARK, a method for online discovery of new categories that leverages the existing feature structure within a closed-set classifier. The authors learn multiple "Local Manifold Anchors" for each known category and construct a shared low-dimensional subspace via SVD. This subspace is then used to constrain the initialization of prototypes for new categories, which are subsequently updated in an online fashion. Experiments conducted on seven public datasets demonstrate that this approach achieves competitive performance on some of these datasets.

Critical strengths
  1. This paper attempts to directly utilize the existing feature structure of closed-set classifiers for online class discovery, resulting in a relatively simple overall approach.

  2. Using local anchors to represent intra-class variations and applying them to initialize new class prototypes is a promising research direction.

  3. This paper provides experimental evidence to motivate the use of local manifold anchors. Figure 1 illustrates the feature structure learned through closed-set supervision, while Table 6 compares different sources for constructing the subspace. The results on CUB are encouraging, suggesting that local manifold anchors may provide a more useful structural feature than class prototypes or instance features.

Critical weaknesses

1. The novelty relative to existing methods needs further clarification.

The core idea of this paper is to use structural information from known classes to assist novel category discovery, but related directions have already been explored in AGE (ICLR 2026). Moreover, PHE (NeurIPS 2024) has already used multiple prototypes to represent intra-class diversity, and Assignment-Driven Hash Learning (CVPR 2026) also constructs multiple prototypes for known classes, so the authors should discuss the differences from these methods. Notably, Eq. (16)–(17) is mathematically similar to the prototype subspace projection mechanism in SAFA-SNN (ICLR 2026), as both combine the original feature with its projection onto a subspace defined by known classes. Although the two methods construct their subspaces differently, the authors should discuss this related work and further clarify the differences in their mathematical formulation and technical contributions.

2. The local specialization of multiple anchors is not sufficiently validated. The authors claim that different Local Manifold Anchors capture distinct local semantic variations. However, Eq. (5) requires each sample to align with more than half of the anchors within its class, meaning that the positive anchor sets of different samples must overlap. I am concerned whether this design can effectively encourage meaningful local specialization, rather than producing largely overlapping anchor representations. Although Fig. 4 studies the sensitivity to (m), the paper lacks direct evidence that the learned anchors actually capture distinct local semantic patterns.

Questions for authors

1. The paper argues that standard closed-set training already preserves useful semantic structures. However, SPARK introduces additional alignment, diversity, and entropy objectives to learn the Local Manifold Anchors. How do the authors distinguish the structural information inherited from the original closed-set features from the geometry introduced by these additional learning objectives?

2. The paper mainly evaluates the Local Manifold Anchors through downstream accuracy. However, it remains unclear whether different anchors actually capture distinct local semantic patterns, as claimed by the authors. Table 3 shows that the diversity and entropy objectives improve performance when used together, but this does not directly demonstrate that the learned anchors represent different local modes. Could the authors provide a more direct analysis of anchor assignments or the samples associated with different anchors to verify their claimed specialization? At the very least, analyses such as activation visualization or correlation analysis are necessary.

3. Eq. (16)–(17) has a similar mathematical form to the prototype subspace projection and weighted fusion mechanism in SAFA-SNN (ICLR 2026). Although the two methods construct their subspaces differently, could the authors clarify the key technical differences and explain what SPARK contributes beyond changing the source of the projection subspace?

Reviewer_Ct73OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPARK, a lightweight framework for On-the-Fly Category Discovery (OCD) that exploits structural knowledge from closed-set classifiers without introducing additional discovery-specific backbone training. It learns Local Manifold Anchors to capture intra-class variations and approximate a shared superclass manifold, which guides analytical initialization and online updating of novel-class prototypes. Experiments on seven benchmarks demonstrate competitive or improved recognition performance over existing OCD methods, supported by ablation studies validating the proposed components.

Critical strengths
  1. The paper explores how structural priors learned by closed-set classifiers can support online category discovery. SPARK avoids additional discovery-specific backbone training, offering a relatively simple and practical solution.
  2. The proposed Local Manifold Anchors capture intra-class variations and guide novel-class prototype initialization. Experiments on seven benchmarks demonstrate competitive performance, while ablation studies support the effectiveness of the proposed components.
Critical weaknesses
  1. The experiments omit recent competitive approaches, such as TALON (CVPR 2026). Comparisons with these methods are necessary to better establish the effectiveness and competitiveness of SPARK.
  2. The proposed framework mainly combines multi-anchor representations, subspace projection, and online prototype updating. The authors should better distinguish their contributions from existing prototype-based and manifold-learning approaches.
  3. The experiments lack detailed evaluations of sensitivity to sample ordering, threshold selection, and error accumulation during online discovery. Such analyses are important for assessing the reliability of SPARK in realistic streaming scenarios.
Questions for authors
  1. Clarify why recent competitive methods are not included in the experiments and discuss how SPARK compares with these approaches.
  2. Further distinguish Local Manifold Anchors and manifold-constrained prototype initialization from existing multi-prototype and subspace-based methods, highlighting their unique contributions.
  3. Explain how sensitive SPARK is to sample ordering, threshold selection, and accumulated assignment errors, and discuss whether the current experimental results support stable long-term online discovery.
Reviewer_4CUEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPARK learns several anchors for each known class on top of a closed-set classifier, without sending the anchor-loss gradients into the backbone. A PCA subspace of these anchors guides the initialization of novel prototypes, which are then updated as test samples arrive. Evaluation covers seven main datasets and four iNaturalist subsets, with additional seed and timing results in the supplement.

Critical strengths
  1. Keeping the closed-set representation intact is a sensible design choice here. Table 5 provides a clear check: stopping anchor gradients raises CUB All accuracy from 35.2% to 45.7%. The prototype initialization is also simple enough to understand and implement.
  2. The supplement answers several practical questions: thresholds come from labeled training data, samples are processed once, and three-seed and full-stream timing results are provided. The method improves All accuracy over AGE on six of the seven main datasets.
Critical weaknesses
  1. My main concern is how well the method actually forms categories. Supplementary Table 14 produces 4,709 categories on Herbarium19, where there are 683, and 1,194 on Cars, where there are 196. This degree of fragmentation matters for online discovery even if clustering accuracy remains competitive. The supplement describes this as evaluation without a known category count, but I could not establish how the main tables use that count or why their protocol differs.
  2. The evidence for a shared superclass manifold is harder to interpret than the method itself. The implemented prior is a global affine PCA subspace. In Table 6, it is unclear whether changing the subspace source also changes recognition or rejection, so the large CUB gain cannot yet be assigned to better initialization. The CIFAR100 result is mixed as well: class-prototype subspaces achieve 40.1% New accuracy, compared with 37.8% for anchors.
Questions for authors
  1. What changes between Tables 1–2 and supplementary Table 14? In particular, is the true category count used for a cap or post-processing, and do the AGE comparisons receive the same information?
  2. Does Table 6 hold recognition anchors, rejection thresholds, and online updates fixed, changing only the initialization subspace? Please clarify this control.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.