acceptodds
Withdrawn from ICLR 2027

Deep Siamese Generative Adversarial Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep clustering has attracted considerable attention owing to the inspiring performance in mining patterns of data. However, previous deep clustering methods cannot capture the category-invariant structures of data in mining patterns. Moreover, they cannot take into consideration the consistencies of data manifolds in learning clustering dependence between categories and instances. To address those challenges, deep Siamese generative adversarial clustering is proposed, which can capture the category-invariant structures and consistent manifolds of data in pattern mining. Specifically, the generative clustering is defined as the generation process of pairwise data within an expectation-maximization scheme, which explores the link information of pairwise data to facilitate consistent structures in clusters and different structures between clusters in fitting data distributions. Furthermore, a deep Siamese generative adversarial clustering network is proposed to implement the definition within a symmetrical architecture that contains generator, discriminator, and cluster networks, which can capture the category-invariant structures and consistencies of data manifolds in clustering. Afterwards, an adversarial clustering loss is defined based on the adversarial game of data distribution and the binary classification task of cluster structures to train the network parameters, which can facilitate the intra-cluster compactness and inter-cluster separation in partitioning data. Finally, extensive experiments are conducted on six benchmark datasets and the results illustrate the state-of-the-art performance of the proposed method in comparison with sixteen cutting-edge methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Deep Siamese generative adversarial clustering, deep expectation maximization, adversarial clustering loss

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Zd6AOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This work introduces Deep Siamese Generative Adversarial Clustering (DSGAC), which casts clustering as a pairwise‑data generation process from category variables to instance variables within an expectation‑maximization (EM) framework. Its E‑step encodes instance‑to‑instance relationships using a category‑aware conditional joint probability matrix, and the M‑step fits data distributions via adversarial training.

Critical strengths

(1)This paper reads well and organize in a good shape.

Critical weaknesses

(1)The combination of GAN and GMM-based clustering is not novel. ClusterGAN [AAAI 2019] or (“Mixture of GANs for Clustering, IJCAI”) already explored latent space clustering with GMM sampling. The Siamese architecture and adversarial clustering loss in this paper represent incremental engineering adaptations rather than a fundamental conceptual contribution. The paper does not sufficiently articulate how its approach differs in kind (not just degree) from prior generative clustering methods. The novelty of this paper is far from the accept bar.

(2)The largest dataset used is CIFAR-100 (60K samples). The paper lacks evaluation on ImageNet-1K or larger-scale benchmarks, which are increasingly standard for demonstrating scalability and practical relevance in deep clustering. This is a significant limitation given that the paper claims “superiority” over state-of-the-art methods.

(3)Several recent and directly relevant methods are missing, including TAC (ICML 2024), CLUDI (ICML 2025) and more (listed bellow). These represent different but important technical directions (external guidance and diffusion-based clustering, respectively). The comparison set is skewed toward older baselines, which weakens the empirical claims.

[1]Universal Guideline-Driven Image Clustering via a Hybrid LLM Agent, CVPR 2026

[2]Semantic-Augmented Image Clustering via Adaptive Multi-Modal Collaboration, AAAI 2026

[3]Parameter-Free Clustering via Self-Supervised Consensus Maximization, AAAI 2026

[4]DeepDP-TGMM: Amortized Non- Parametric Clustering for Hyperspherical Self-Supervised Representations, CVPR 2026

(4)The paper does not address the well-known training instability associated with GANs in clustering tasks. No convergence curves, variance across runs, or stability analysis are provided. The reported stable accuracy is therefore difficult to assess. I recommend the authors provide multiple-run statistics (mean ± std) and discuss how they mitigate GAN training instability.

Questions for authors

Please refer to the weakness.

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

This paper proposes Deep Siamese Generative Adversarial Clustering (DSGAC), which integrates adversarial learning with pairwise clustering constraints by generating sample pairs with must-link and cannot-link relationships. The method enables end-to-end clustering representation learning.

Critical strengths
  1. The integration of adversarial learning and pairwise clustering constraints offers a certain degree of novelty.
  2. The proposed framework jointly optimizes the generator, discriminator, and clustering network in an end-to-end manner, providing a well-structured architecture.
  3. The method achieves competitive results on six benchmark datasets in terms of ACC, NMI, and ARI, demonstrating its empirical effectiveness.
Critical weaknesses
  1. Issues with theoretical justification. The logarithmic GAN objective in Eq. (5) is replaced by a non-logarithmic objective in Eq. (11) without explanation. Moreover, Eq. (12) directly interprets cosine similarity as the probability of two samples belonging to the same cluster, which is inconsistent with Eq. (4). Under the independence assumption, this probability should be the dot product of the two assignment vectors. For example, when both vectors are uniform, their cosine similarity is 1, whereas the same-cluster probability is 1/K.
  2. Insufficient justification of the core mechanism. The clustering loss in Eq. (13) is defined solely on generated sample pairs, while clustering inference in Eq. (14) is performed on real data. The paper does not adequately explain how the learned pairwise constraints generalize to real samples. Furthermore, the adversarial objective in Eq. (11) primarily matches the marginal distributions of real and generated data, without explicitly ensuring consistency between their class-conditional distributions.
  3. Insufficient experimental validation. The paper lacks ablation studies to assess the individual contributions of the Siamese architecture, adversarial learning, and pairwise constraints. Additionally, although the results are averaged over five independent runs, standard deviations are not reported.
Questions for authors
  1. Could the authors provide a rigorous derivation of the EM formulation and clarify the relationship between Eqs. (5), (11), (12), and the actual implementation?
  2. Could the authors provide ablation studies of the key components and explain how pairwise constraints learned from generated samples improve clustering performance on real data?
Reviewer_gkE4OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, a deep Siamese generative adversarial clustering method is proposed to capture the category-invariant structures and consistent manifolds of data for pattern mining. In detail, generative clustering is defined as the generation process of pairwise data to explore the link information of pairwise data, which facilitates consistent intra-cluster and inter-cluster structures in fitting data distributions. A deep Siamese generative adversarial clustering network along with an adversarial clustering loss is defined to implement the generation process of pairwise data, which learns clustering patterns in an end-to-end manner. Finally, extensive experiments are conducted on 6 data sets, and the results illustrate the state-of-the-art performance.

Critical strengths

S1: A new clustering method is introduced via defining the generation process of pairwise data from category variables to instance variables, which explores the links between instances to capture consistent structures in clusters and different structures between clusters in fitting data distributions. S2: A deep Siamese generative adversarial clustering network is devised to learn clustering-specific representations and partition data in an end-to-end manner, which can capture the category-invariant structures and the consistencies of data manifolds. S3: Extensive experiments are conducted on 6 benchmark datasets, illustrating the superiority of DSGAC compared with state-of-the-art baseline methods in terms of quantitative and qualitative results.

Critical weaknesses

W1: The generation process of pairwise data from category variables for clustering is very interesting. In the manuscript, the advantages of the new generative clustering could be more clearly stated. W2: A deep Siamese generative adversarial clustering network is proposed to implement generative clustering. The motivations of the Siamese architecture could be further illustrated in the manuscript. W3: The quantitative and qualitative results verify superior performance compared with state-of-the-art baseline methods. The core supervision for clustering derives from generated pairs with synthetic links, and the real-data alignment loss could be more clearly described in the manuscript.

Questions for authors

Q1: How is the cluster network exposed to real data during training? Is there any loss that directly aligns indicator features for real samples, or is all cluster supervision derived from generated pairs only? Q2: Please clarify the discriminator design: are the two Siamese discriminators weight-shared or independent? Why are two discriminators needed instead of a single shared one? Q3: Equation (5) uses a log-based GAN objective while Eq. (11) omits logs; what is the rationale for this change, and how does it affect the EM interpretation and convergence properties.

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

Related papers

Discussion (0)

Sign in to comment.