acceptodds
Under review as a conference paper at ICLR 2027

Discovering or Prescribing? Reflexive Structural Supervision for Contrastive Clustering

Abstract

Self-supervised learning seeks to derive learning signals from unlabeled data itself, allowing meaningful representations to emerge without externally provided semantic supervision. For clustering, however, this principle leads to a fundamental tension. The organization of the data is precisely what the model is expected to discover, yet many contrastive clustering methods construct their supervision under an assumed clustering structure before that organization has been revealed. This tension is particularly relevant in unlabeled data, where multiple levels of organization may coexist and the appropriate granularity is generally unknown a priori. Committing representation learning to a single prescribed view of the data may therefore favor one possible organization while leaving other meaningful structural relationships under-explored. We develop Discovering Intrinsic Structure via Contrastive Clustering (DISC) as a reflexive learning framework, in which structural supervision is progressively inferred from the current representation space and fed back into representation learning. To avoid overlooking complementary structural information under a single prescribed view, DISC derives supervisory signals from discovered multi-granularity structural relationships. Furthermore, to reduce discrepancies between assignment results across granularities, DISC imposes cross-granularity consistency constraints between adjacent levels. Comparisons with recent alternatives demonstrate that DISC achieves strong representation quality and clustering performance without relying on a predefined clustering structure, while the ablation results consistently show the benefit of discovering and exploiting multi-granularity structure throughout learning. These findings suggest a shift in contrastive clustering from learning under a prescribed structure toward learning supervision through progressive structure discovery.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Deep Clustering, Contrastive Clustering, Self-Supervised 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_1EJHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DISC, a contrastive clustering framework motivated by the observation that many existing methods construct self-supervised signals using a predefined clustering structure, even though discovering that structure is itself the objective of clustering. DISC instead alternates between structure discovery and representation optimization. An EMA teacher extracts representations from which FINCH discovers a multi-granularity hierarchy, while a student encoder is optimized using a unified instance/prototype contrastive objective and a cross-granularity alignment objective. The resulting representation is subsequently used for another round of structure discovery.

Experiments on several clustering benchmarks report competitive or strong clustering performance. Ablation studies investigate the effects of multi-granularity supervision, hierarchical relationships, repeated structure discovery, the momentum teacher, and the two proposed loss components.

Critical strengths
  1. A compelling formulation of multi-granularity structural supervision. Rather than committing representation learning to a single predefined clustering granularity, DISC repeatedly discovers a hierarchy from the evolving representation and uses multiple granularities as supervision. This is a natural approach to the unknown-granularity setting and gives the method a clear conceptual motivation.

  2. Strong empirical results accompanied by useful ablations. DISC reports strong performance across several benchmark datasets, and the ablations provide evidence that the main components are consequential: hierarchical multi-granularity supervision outperforms single or independent multi-granularity supervision, while repeated structure discovery, the momentum teacher, and the alignment objective generally improve over their corresponding variants.

Critical weaknesses

The evaluation of the discovered hierarchy is insufficiently clear. DISC produces multiple clustering granularities, but the main tables report a single ACC/NMI/Purity/F-score per dataset. Importantly, the discovered hierarchy does not necessarily contain a level corresponding to the ground-truth class count; for example, the reported CIFAR-10 hierarchy has 1974, 393, 74, 16, 4, and 2 clusters. It is therefore unclear to me how the particular clustering used to compute the headline metrics is selected. This is consequential to the central claim that DISC does not need a predefined clustering structure: if ground-truth labels, ground-truth K, or retrospective selection of the best hierarchy level enters this evaluation, the interpretation of the results changes substantially.

Questions for authors
  1. How is the single clustering partition used to report ACC/NMI/Purity/F-score selected from DISC's final multi-granularity FINCH hierarchy? For example, the reported CIFAR-10 hierarchy contains K = 1974, 393, 74, 16, 4, and 2, rather than K = 10. Does the selection procedure use ground-truth K, labels, or any other dataset-specific information? Please describe the complete selection/evaluation procedure. This clarification could materially affect my assessment of the parameter-free/unknown-structure claim.
  2. The main text defines the alignment loss using KL divergence between the projected child distribution and the parent distribution, while the expression in Appendix B.2 appears to place the parent distribution in the numerator of the KL expression. Since KL divergence is asymmetric, could the authors clarify which direction is actually implemented and reconcile the equations?
2 comments

W1 & Q1: The Complete Evaluation Procedure

To evaluate clustering performance, we apply K-Means to the feature representations learned by DISC, using the ground-truth number of classes KK to specify the number of clusters. The resulting single clustering partition is used to compute the reported clustering metrics. This evaluation protocol follows common practice in deep clustering.

The parameter-free property of DISC is primarily reflected in its use of FINCH to automatically discover clustering structures based on nearest-neighbor relations. It progressively constructs a multi-granularity hierarchy through iterative merging, without requiring any predefined clustering-related parameters. This mechanism enables DISC to jointly perform structure discovery and feature representation optimization under unknown clustering structures.

We have clarified the evaluation protocol (Lines 354–359) and provided a brief description of FINCH in the revised manuscript (Lines 236–238).

Q2: KL Divergence Direction in the Multi-Granularity Alignment Loss

The KL divergence implemented in our code takes the projected child-level assignment distribution as the first argument and the corresponding parent-level assignment distribution as the second, consistent with the definition in Equation (8) of the main text. This alignment loss promotes consistency in the multi-granularity hierarchy by constraining the assignment distributions across adjacent hierarchy levels. In response to the reviewer's comment, we have revised the loss expression and its gradient derivation in Appendix B.2 to ensure consistency with Equation (8).

AuthorsOpenReview ↗

Follow up on our responses

Dear Reviewer 1EJH,

Thank you for your previous valuable comments. In our responses, we have addressed your concerns by clarifying the evaluation pipeline and multi-granularity hierarchical structure discovery mechanism of DISC (W1 & Q1), as well as checking the consistency of the implementation with Equation (8) and correcting the relevant expressions and derivations (Q2).

We would sincerely appreciate it if you could review our responses when convenient and let us know whether these explanations have addressed your concerns. We would be happy to discuss any remaining questions or provide further clarification.

Thank you very much for your time and consideration.

Reviewer_YfoLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces Discovering Intrinsic Structure via Contrastive Clustering (DISC), a contrastive clustering method. DISC is founded on the observation that many existing methods implicitly or explicitly impose a predefined cluster structure on the data, for example through priors on the cluster number or cluster size. These priors may not be available in real applications, or the may not fit the data, so it is argued that clustering methods should be free of such priors. DISC is proposed as a prior-free clustering method based on contrastive learning. After training an autoencoder to obtain initial sample features, EM-style optimization is applied, where FINCH is used to obtain a hierarchical clustering of the data during the E-step, and contrastive representation learning is used to align the features with the clustering during the M-step. Experiments show DISC outperforming existing methods on most datasets, with an ablation study confirming the utility of the individual components.

Critical strengths
  • The aim of avoiding to force a certain cluster structure is sound and well-motivated.
  • The proposed method outperforms the baselines in the supplied experiments, in some cases substantially. The provided ablation study allows detailed insight into the components that facilitate the gains.
Critical weaknesses
  • The method's novelty is limited, in essence DISC is PCL with FINCH replacing k-means as the clustering method, with a consistency loss to ensure the features reflect the hierarchy.
  • I have several concerns regarding the evaluation. First, the results on CIFAR-10/20 and STL-10 appear to compare different models, the baseline numbers are taken directly from the MiniClustering paper and use ResNet 18, while DISC uses an MLP autoencoder architecture. If this is the case then the comparison in Table 2 is invalid and must be redone. Second, the paper states that "FINCH is applied to the final Teacher representations to obtain the multi-granularity clustering results", but the supplied code uses standard k-means with the ground-truth cluster number to generate the final predictions. This mismatch requires an explanation, clearly either the paper text or the supplied code are incorrect. Furthermore, the decision to only train on a subset of MNIST, CIFAR, and Fashion-MNIST is unexplained and may impact results compared to training on the entire dataset. Finally, the results would be more convincing if they also included more complex model architectures that approach what practitioners may use, the MLP autoencoder used in the paper is far from that.
Questions for authors
  • How well do the discovered hierarchies correspond to ground truth ones on hierarchical datasets?
8 comments

W1: Regarding the Novelty of the Proposed Method

DISC extends contrastive clustering by tightly integrating multi-granularity hierarchical structure discovery with feature representation optimization. It challenges the field's reliance on predefined prior information and offers a new perspective for advancing contrastive clustering toward truly autonomous structure discovery. Its novelty is reflected in three aspects:

  1. From human-specified structures to autonomous discovery of intrinsic data organization. Existing methods typically introduce prior assumptions about data organization into the model, limiting their applicability in open environments. In contrast, DISC treats the structure itself as an object to be explored during learning, enabling the model to discover unknown clustering structures directly from the data and move beyond reliance on predefined prior information.

  2. From single-granularity structural understanding to complementary integration of local and global information. Existing methods are often restricted to a single granularity, making it difficult to capture the full complexity of data organization. DISC integrates fine-grained local sample relationships and coarse-grained global semantic associations into a unified contrastive learning framework, allowing structural information at different levels to jointly contribute to supervision.

  3. Iterative co-evolution of hierarchical structures and feature representations. To the best of our knowledge, DISC is the first contrastive clustering framework to achieve the co-evolution of multi-granularity hierarchical structures and feature representations through alternating optimization, without relying on predefined prior information.

W2.1: Model Architecture and Baseline Comparability

In the experiments on CIFAR-10/20 and STL-10, both DISC and MiniClustering use features extracted by ResNet-18 as inputs for subsequent clustering optimization, ensuring a fair comparison based on the same feature representations. This concern may have arisen because the supplementary code mainly presents the overall architecture of DISC without specifying the settings used in the experiments on imbalanced datasets. We have clarified the experimental settings for imbalanced datasets in the manuscript (Lines 1076–1078).

W2.2: Consistency Between the Paper Description and Code Implementation

Both the paper description and the code implementation are functionally consistent. K-Means is used solely for external evaluation to assess the clustering performance of the feature representations refined through multi-granularity structural guidance in DISC. This is also a commonly adopted evaluation practice in deep clustering.

FINCH is applied to the final Teacher representations to obtain the complete multi-granularity hierarchy, which represents the overall structural discovery result of DISC.

We have clarified the evaluation protocol in the manuscript accordingly (Lines 354–359).

W2.3: Evaluation on Full Datasets

In our initial experiments, we used subsets of MNIST, Cifar10, and Fashion-MNIST, primarily because some baseline methods incur high computational costs when applied to the full datasets. This allowed us to compare different methods within a manageable computational budget.

To further address this concern, we conducted experiments using the complete training sets of these three datasets. The detailed results are presented in the table below.

Dataset#SamplesACCNMIPURF-score
MNIST600000.78550.77980.78550.7822
Fashion-MNIST600000.60140.64280.62950.5775
Cifar10500000.89820.78870.89820.8977

We have added the corresponding explanation to the manuscript, and the detailed experimental results can be found in Appendix G.

Q1: Alignment Between Discovered and Ground-Truth Hierarchical Structures

DISC discovers multi-granularity hierarchical structures that exhibit a high degree of consistency with the ground-truth category organization. To further investigate this, we conducted additional visualization experiments on the discovered hierarchical structures. Specifically, we visualized the hierarchical clustering results across six datasets, with clusters color-coded according to their corresponding ground-truth categories and cluster purity.

We observe three notable patterns:

  1. Category separation across granularities. At each granularity, clusters of different colors form relatively distinct and coherent color blocks, indicating clear category separation.
  2. Consistency across granularities. Coarse-grained clusters and their corresponding fine-grained sub-clusters often share the same color, indicating consistency in category assignments across hierarchical levels.
  3. Consistency among fine-grained sub-clusters. Multiple fine-grained sub-clusters under the same coarse-grained cluster also typically exhibit the same color, suggesting that they predominantly belong to the same ground-truth category.

These observations suggest that the hierarchical structures discovered by DISC are highly consistent with the ground-truth category organization and can further reveal fine-grained structures within individual categories. The corresponding visualizations and analysis have been added to Appendix F of the manuscript.

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

The paper proposes DISC, a contrastive clustering framework that jointly refines representations and discovers hierarchical groupings without predefined cluster counts. Its most appealing aspect is adaptive granularity: the scales used to supervise learning can evolve with the representation. The combination of hierarchical prototype supervision, cross-level consistency, and a momentum teacher forms a coherent framework. The reported clustering results are encouraging, and the ablations support the usefulness of repeated structure discovery and cross-level alignment. The presentation could be strengthened with qualitative examples showing the semantic relationships captured at different levels and plots illustrating how cluster counts evolve during training. Additional runtime measurements and practical guidance on parameter selection would also improve clarity and usability. These additions would help readers interpret the discovered structures and assess the framework’s practical behavior. Overall, the paper addresses a meaningful problem and offers a useful extension of clustering-based representation learning. Given its coherent methodology and encouraging empirical evidence, I recommend weak accept.

Critical strengths
  1. Adaptive granularity addresses a meaningful limitation. Rather than prescribing several cluster counts, DISC allows both the groupings and their granularities to evolve with the representation. This is a useful extension of methods such as PCL, particularly when the appropriate clustering resolution is unknown.

  2. The components form a coherent learning mechanism. FINCH supplies a hierarchy, prototype contrastive learning converts it into supervision, cross-level alignment coordinates the granularities, and the momentum teacher stabilizes the features used for subsequent discovery.

  3. The experiments provide encouraging evidence. DISC reports strong clustering performance across several datasets. The repeated-versus-one-shot discovery ablation is particularly informative: On MNIST, full DISC achieves 83.36% ACC, compared with 59.86% for the one-shot structure-discovery variant.

Critical weaknesses
  1. More qualitative examples would improve interpretability. Representative samples from fine and coarse clusters would help readers understand what relationships the discovered hierarchy captures and how the granularities complement one another.

  2. The evolution of the hierarchy could be illustrated more clearly. A simple plot of cluster counts across training rounds would make the adaptive-granularity mechanism more tangible and show when the discovered organization becomes stable.

  3. Practical parameter guidance could be expanded. A brief discussion of how to choose the teacher momentum, hierarchy-update frequency, and loss weights would make the framework easier to reproduce and apply to new datasets.

  4. Computational overhead could be reported more concretely. Although the paper discusses computational complexity, a small runtime breakdown separating hierarchy discovery from representation optimization would help readers assess the practical cost of repeatedly updating the clustering structure.

Questions for authors
  1. How do the cluster counts and hierarchy depth evolve across training rounds? Does DISC typically discover finer or coarser partitions as the representation improves, and when do these granularities become stable?

  2. Could you provide representative examples from different hierarchy levels to illustrate the relationships captured by the fine and coarse clusters?

  3. What proportion of training time is spent on FINCH hierarchy discovery? How did you choose the hierarchy-update frequency, and what trade-off have you observed between computational cost and clustering performance?

3 comments

W1 & Q2: Visualization of the Discovered Hierarchical Structures

The hierarchical structures discovered by DISC reveal nested subcategories within the same category, where samples belonging to a coarse-grained category can be further divided into multiple fine-grained sub-clusters with more specific characteristics. Different granularities characterize the category organization at different levels: coarse-grained clusters provide an overall category-level partition, while fine-grained clusters further distinguish intra-category variations, thereby capturing more detailed local structures. To illustrate these relationships, we conducted additional visualization experiments on the hierarchical clustering results across six datasets, with clusters color-coded according to their corresponding ground-truth categories and cluster purity.

We observe the following patterns:

  1. Nested structures within categories. Multiple fine-grained sub-clusters under the same coarse-grained cluster typically exhibit the same color, indicating that these sub-clusters predominantly belong to the same ground-truth category and reveal its internal substructures.
  2. Complementarity across granularities. Coarse-grained clusters and their corresponding fine-grained sub-clusters often share the same color, indicating that fine-grained partitioning preserves the category semantics of coarse-grained clusters while further revealing more detailed structures within each category.
  3. Category separation across granularities. At each granularity, clusters of different colors generally form relatively distinct and coherent color blocks, indicating that DISC maintains clear category separation while progressively refining the clustering structures.

These observations suggest that DISC not only discovers nested substructures within categories but also captures both the overall category organization and fine-grained intra-category variations through complementary information across granularities. The corresponding visualizations and analysis have been added to Appendix F of the manuscript.

W2 & Q1: Evolution of the Discovered Hierarchical Structures

In the early stages of training, the cluster counts across different hierarchy levels and the overall hierarchy depth exhibit noticeable fluctuations. As the feature representations are progressively refined, these fluctuations gradually diminish, and the hierarchy becomes increasingly stable in the later stages. DISC does not consistently favor finer or coarser partitions but adaptively adjusts the clustering granularity according to changes in the intrinsic data organization.

To address the reviewer's suggestion for visualizing the evolution of the discovered hierarchy, we recorded the cluster counts at different hierarchy levels across EM iterations on three datasets and plotted their evolution to illustrate the dynamic changes in the hierarchical structure.

As shown in the plots, the curves of cluster counts at different granularities become generally smoother in the later stages than in the early stages across HW2, Wiki, and MNIST, indicating that the clustering structures at different granularities gradually stabilize. The corresponding visualizations have been added to Appendix C.2 of the manuscript.

W3: Practical Hyperparameter Settings

In our experiments, the momentum coefficient of the Teacher is set to 0.99, and the hierarchical structure is updated every 10 inner training epochs, with a total of 50 EM iterations. Regarding the loss weights, the weight of the Unified ProtoNCE loss follows an exponential decay schedule from 1 to 0.01, while the weight of the multi-granularity alignment loss is typically set within the range of 0.3–0.5. We have incorporated the corresponding parameter details into the manuscript (Lines 317–322).

W4 & Q3: Computational Overhead and Hierarchy-Update Frequency

We update the hierarchical structure every 10 inner training epochs, with a total of 50 EM iterations. This update frequency aims to balance computational cost and clustering performance: more frequent updates would increase the computational cost of hierarchy discovery, whereas excessively long update intervals might prevent structural supervision from adapting promptly to changes in feature representations, leading to degraded clustering performance.

To further quantify the computational overhead introduced by hierarchical structure discovery, we separately measured the runtime of hierarchy discovery and feature representation optimization. The detailed results are presented in the following table.

DatasetHierarchy Discovery (s)Feature Optimization (s)Total (s)
HW219.20 (32.0%)40.83 (68.0%)60.03
Wiki22.23 (25.0%)66.84 (75.0%)89.07
MNIST55.31 (26.3%)155.00 (73.7%)210.31
USPS135.83 (23.7%)436.75 (76.3%)572.58
Cifar10185.19 (25.3%)545.72 (74.7%)730.91
Fashion-MNIST183.81 (24.1%)577.94 (75.9%)761.75

As shown in the table, hierarchy discovery accounts for only 23.7%–32.0% of the total training time across the six datasets, while feature representation optimization accounts for the majority of the training time. This indicates that although DISC repeatedly performs hierarchical structure discovery during training, the resulting computational overhead remains manageable.

We have added the corresponding computational overhead analysis (Lines 484–485) and explanation of the hierarchy-update frequency to the manuscript (Lines 317–322).

Comments on the paper

5 comments

W1 & Q1: Generation and Evaluation of the Final Clustering Results

To evaluate clustering performance, we apply K-Means to the learned feature representations, using the ground-truth number of clusters. The single-partition results are obtained through this evaluation protocol and do not correspond to any specific hierarchy level.

This protocol is consistent with the evaluation methods employed in most existing deep clustering and multi-granularity structure learning methods, enabling a fair comparison of the learned feature representations.

We have clarified the evaluation protocol in the revised manuscript (Lines 354–359).

W2 & Q2: Relationship to ReSA and Novelty of DISC

Explicit multi-granularity hierarchy discovery and parent-child mapping constraints are among the key methodological differences between DISC and ReSA. More fundamentally, the two methods differ in their research objectives. ReSA focuses on self-supervised representation learning, generating online supervisory signals from the clustering properties inherent in feature representations to improve representation quality. In contrast, DISC is designed for contrastive clustering, aiming to move beyond predefined prior information by autonomously discovering the intrinsic multi-granularity structure of data and incorporating it into the clustering optimization process. We have added the citation to ReSA (Lines 137–141) and revised the manuscript to clarify DISC's contributions (Lines 99–100).

The new paradigm introduced by DISC aims to break the fundamental dependence of existing contrastive clustering methods on predefined prior information. Existing methods often take default structural priors as a prerequisite for their operation. Without such prior information, they may struggle to construct effective training supervision or even determine the network architecture.

To the best of our knowledge, DISC is the first contrastive clustering framework to tightly integrate multi-granularity hierarchical structure discovery with feature representation optimization within an alternating optimization framework. It not only removes the field's naive assumption of relying on predefined prior information, but also offers a new perspective for advancing contrastive clustering toward truly autonomous structure discovery.

Dear Area Chair and Reviewers,

We would like to express our sincere gratitude to the AC for overseeing this paper, and to the reviewers for the valuable feedback.

We submitted our detailed responses to the review comments on November 13. Following your suggestions, we also revised the manuscript accordingly.

As the ICLR 2027 author-reviewer discussion period will end on November 18, 2026, only a few days remain before the discussion deadline. We would greatly appreciate it if you could review our responses and the revised manuscript at your earliest convenience and provide any further feedback, so that we may address any remaining concerns before the discussion period concludes.

Thank you for your time and consideration!

Thank you for the clarification and the specific revisions to the paper. My concerns have been largely addressed, and I have updated my rating accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.