acceptodds
Under review as a conference paper at ICLR 2027

SAAS: Size-Aware Anchor Selection for Generalized Category Discovery

Abstract

Generalized category discovery (GCD) assigns unlabeled images to known and novel categories using supervision from known classes. Reliable groups of similar images can help correct category predictions, but assessing their reliability through graph density can penalize larger groups. This bias arises because sparse nearest-neighbor graphs limit the connections available to each image, whereas complete-graph normalization counts all possible connections within a group. We propose Size-Aware Anchor Selection (SAAS), a general reliability assessment strategy for selecting reliable local groups under sparse graph constraints, instantiated as a post-hoc prediction correction framework without additional training. SAAS normalizes the observed connections by the maximum number permitted by the group size and neighborhood size. It selects groups whose images are similar and whose membership remains consistent across two augmented views, using these groups as anchors for correction. Similarity to the selected groups guides corrections to ambiguous predictions. We evaluate SAAS on six generic and fine-grained GCD benchmarks using DINOv2 pretraining. On CUB, the resulting pipeline achieves 88.67% overall accuracy, with 85.80% on known classes and 90.10% on novel classes. Ablation studies show that removing the group-size cap alone leaves predictions unchanged, whereas the proposed normalization improves accuracy and enables further gains when the cap is removed. These results highlight the importance of accounting for sparse connectivity when selecting reliable groups for GCD prediction correction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
generalized category discovery, anchor selection, neighborhood graphs, graph-based clustering, prediction correction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HbdnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Size-Aware Anchor Selection (SAAS), an extension for k-means-based GCD methods such as HypGCD that aims to correct predictions for samples near the decision boundary. For this, connected components of samples sharing the same provisional label are identified based on a mutual qNN graph. These groups are then filtered for view-consistency and compactness to improve the reliability of the selected groups. In addition, the groups are filtered based on internal connectivity using a formulation that takes into account that the group's connection budget scales linearly with group size. The selected groups are used as anchors to correct the cluster assignments of samples with a small relative distance margin between their two nearest cluster centers. Experiments show SAAS outperforming standard HypGCD on CUB, while achieving similar performance to HypGCD on other benchmarks. Analysis on CUB shows SAAS outperforming alternative correction baselines.

Critical strengths
  • The fundamental idea of identifying samples near the decision boundary and correcting their cluster assignments using anchor groups that exhibit certain consistency criteria is well-motivated. The proposed method achieves an improvement on the CUB dataset.
Critical weaknesses
  • The paper's contribution to the field is fundamentally limited, there are no major theoretical insights, and the empirical results only show a relevant improvement on one dataset. From a practical perspective, this small improvement is not worth the added complexity and the additional hyperparameters. In addition, SAAS requires two views for each image, increasing feature extraction cost compared to the baseline. Given that the contribution is limited from both a theoretical and a practical perspective, the work cannot be accepted.
  • CUB is the only dataset on which SAAS achieves good results, while other datasets only show negligible improvements at best. The reason behind this discrepancy is not investigated in the paper, there is no attempt to identify conditions under which SAAS works well and why. In general a lot of the analysis performed in the paper only considers CUB, leaving the performance of the method poorly understood.
Questions for authors
  • What is the cause for the differences between the Hyp-SelEx results reported here and in the original paper? Especially CUB and Stanford Cars show large differences.
Reviewer_H4kLOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies post-hoc correction of ambiguous predictions in clustering-based Generalized Category Discovery (GCD). Corrections are guided by reliable local "anchor groups," and no retraining is required.

The key observation concerns mutual-qNN graphs, where each node has at most q edges. On such graphs, the usual complete-graph density E/[n(n−1)] implicitly caps group size at n ≤ 1 + q/τ_ρ. The authors therefore replace it with a connection-budget density, E/[n·min(q, n−1)].

The method works as follows:

  • Candidate groups are connected components within each initial cluster. They are filtered by flip-view Jaccard stability, feature compactness, and budget density.
  • A sample is treated as ambiguous when its distances to the two nearest cluster centers are similar. If both candidate classes have anchor groups, the prediction is switched to the second class when that class has stronger mean or max anchor similarity and a cross-view neighborhood check also agrees.

The paper reports the following results:

  • On CUB, applied on top of Hyp-SelEx, All accuracy goes from 86.65 to 88.67 (+2.02 pp).
  • On the other five datasets, paired changes range from −0.02 to +0.39 pp.
Critical strengths

The paper identifies a genuine mismatch between complete-graph density and the bounded degree of mutual-qNN graphs, and Proposition 1 formalizes it correctly. The resulting correction step requires no retraining and adds little computational overhead.

Critical weaknesses
  1. The headline results are arithmetically inconsistent with the paper's own correction counts

Table 1 reports a CUB gain of +2.02 pp (86.65 to 88.67), yet Table 3, which covers the same 4,496 images over three seeds, reports only 19 modified predictions for the mean rule and 15 for the max rule; even if all 19 changes fell in one seed and were all fixes, the gain could be at most about 0.42 pp. The most plausible explanation is that the baseline and SAAS rows of Table 1 come from different clustering runs, which would make the headline number an unpaired comparison rather than a gain attributable to SAAS. The same inconsistency recurs elsewhere. Section 4.2 reports identical CUB accuracies for the mean and max variants to four decimals (88.6655 / 85.8039 / 90.0968), although Table 3 shows that they modify different numbers of predictions (19 vs. 15) with different fix counts (13 vs. 9), and Tables 9 and 11 in the appendix repeat the pattern. On CIFAR-100, Section 4.2 reports a paired gain of +0.08 pp, whereas Table 9 shows that each run modifies only 14 (mean) or 12 (max) of 30,000 predictions with net changes of +3 and +5, which bounds the gain at about 0.047 pp even if the Hungarian matching is recomputed. Finally, Appendix E describes Table 10 as computed "before additional neighborhood confirmation", yet it reports exactly the gains of Table 2 (+0.1093 mean, +0.1553 max); either the confirmation step of Section 3.4 has no effect, or the protocol is misdescribed.

  1. The contribution is a one-line change of normalization, and alternatives are not compared

The core change is the density denominator, from n(n−1) to n·min(q, n−1). The rest of the framework is inherited: mutual-qNN graphs (Maier et al., 2009), Jaccard stability (Hennig, 2007), and the anchor-based correction procedure itself, which the paper calls "the existing correction setup" without a citation. If that procedure is prior work, the contribution reduces to the new denominator, and the paper does not compare it with other size-robust reliability measures such as average internal degree, k-core or conductance.

Questions for authors
  1. How is the +2.02 pp CUB gain in Table 1 obtained, given that Table 3 shows only 19 modified predictions? Do the baseline and SAAS rows come from the same clustering outputs, and how many seeds are used? Please report the paired ΔAll for CUB, as Table 11 does for the other datasets.
  2. Why do the mean and max variants give identical accuracies (Section 4.2, Table 11) but different correction counts (Tables 3 and 9)? How does the +0.08 pp gain on CIFAR-100 reconcile with the counts in Table 9?
  3. Does Table 10 exclude the cross-view confirmation step, as Appendix E suggests? If so, why are its gains identical to Table 2? Please define the confirmation rule and state whether it also applies to the KNN and nearest-member baselines.
  4. Section 4.3 uses Table 3 to argue that SAAS corrects more precisely than KNN. However, on the development split used for Table 2, KNN reaches 84.62 while the best SAAS variant reaches 84.58, as shown in the appendix. On CIFAR-100, KNN achieves a net correction of +33 to +36 per run, compared with +3 or +5 for SAAS. How should these results be interpreted relative to the claimed advantage of SAAS?
Reviewer_6wxPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAAS selects reliable groups in sparse neighbor graphs to correct ambiguous GCD predictions. Its main idea is a size-aware connectivity normalization that avoids penalizing larger groups.

Critical strengths
  1. The motivation is clear, and the method requires no additional training, which is good.
  2. The ablation usefully separates normalization from the group-size cap.
Critical weaknesses
  1. The method is straightforward but seems not effective. Experiments show that the improvement is limited.
  2. Table 1 reports a ~2.0 improvement on CUB, but Tab 3 reports only 19 modifications across 13K events. For the same paired predictions, these could improve accuracy by at most 0.1409 points.
Questions for authors
  1. Could the authors check the results in Tables 1 and 3?
  2. Given the common-support results in Table 3, what additional benefit does group-based scoring provide over simple nearest-member correction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.