Source-linked AI summary
Deep Comprehensive Correlation Mining for Image Clustering
Jianlong Wu, Keyu Long, Fei Wang, Chen Qian, Cheng Li, Zhouchen Lin, Hongbin Zha
TL;DR
Deep clustering methods often focus on pair-wise sample correlations, leaving category information and other useful correlations underused. DCCM progressively combines pseudo-label supervision, transformation robustness, and triplet mutual information to learn discriminative representations, achieving significant improvement over state-of-the-art methods, including 0.623 clustering ACC on CIFAR-10, 10.1% above DAC.
Problem
Existing deep clustering methods mainly focus on pair-wise sample correlations, while category information and other useful correlations remain underused.
Method
DCCM progressively mines highly confident pseudo-labels, transformation robustness, and triplet-level mutual information to learn discriminative features.
Results
DCCM significantly surpasses state-of-the-art methods across the evaluated benchmarks and metrics, reaching 0.623 clustering ACC on CIFAR-10, 10.1% higher than DAC.
Takeaways & Limitations
Mining multiple correlations beyond sample pairs supports progressive unsupervised training and improves deep clustering performance.
Takeaways & Limitations
The local robustness formulation assumes that geometry-transformed images retain the same label and similar features.
Abstract
from arXiv · showhide
Recent developed deep unsupervised methods allow us to jointly learn representation and cluster unlabelled data. These deep clustering methods mainly focus on the correlation among samples, e.g., selecting high precision pairs to gradually tune the feature representation, which neglects other useful correlations. In this paper, we propose a novel clustering framework, named deep comprehensive correlation mining(DCCM), for exploring and taking full advantage of various kinds of correlations behind the unlabeled data from three aspects: 1) Instead of only using pair-wise information, pseudo-label supervision is proposed to investigate category information and learn discriminative features. 2) The features' robustness to image transformation of input space is fully explored, which benefits the network learning and significantly improves the performance. 3) The triplet mutual information among features is presented for clustering problem to lift the recently discovered instance-level deep mutual information to a triplet-level formation, which further helps to learn more discriminative features. Extensive experiments on several challenging datasets show that our method achieves good performance, e.g., attaining $62.3\%$ clustering accuracy on CIFAR-10, which is $10.1\%$ higher than the state-of-the-art results.
1. Introduction
DCCM addresses limitations of pair-wise deep clustering by progressively mining pseudo-label, transformation-robustness, feature-level, and inter-correlation information for discriminative representations.
- Motivation: Traditional image clustering struggles with real-world variation, limited hand-crafted features, and separating feature extraction from clustering.Deep unsupervised methods jointly learn representations and clustering to address these limitations.
- Motivation: Existing methods often lack discriminative power and DAC mainly uses pair-wise correlations, neglecting category information and other useful correlations.The paper identifies pseudo-label, transformation, and feature correspondence information as additional sources of supervision.
- DCCM Framework: DCCM is an end-to-end framework that progressively mines highly confident information from multiple correlations.It considers correlations among samples, local robustness to geometry transformations, corresponding features across layers, and their inter-correlations.
- DCCM Framework: Pseudo-label supervision uses highly confident predictions to investigate category information and guide unsupervised network training.The pseudo-label loss complements pair-wise pseudo-graph supervision.
- DCCM Framework: Local robustness exploits transformed inputs, while triplet mutual information extends instance-level mutual information to learn more discriminative features.These mechanisms use feature correspondence and image-transformation robustness beyond pair-wise sample correlations.
2. Related Work
Prior work combines deep feature learning with clustering objectives, designs specialized representation losses, and uses self-supervised pretext tasks to learn semantic features without labels.
- Deep Clustering: Deep clustering methods combine deep feature learning with traditional clustering objectives, including KL-divergence, K-means, and subspace-clustering losses.Autoencoders are a common basis for learning latent representations.
- Deep Clustering: Other approaches design losses directly on final-layer outputs, merge nearby clusters, use sample correlations as supervision, or extend spectral clustering.These methods differ in how they impose clustering structure on learned features.
- Unsupervised Representation Learning: Deep unsupervised representation learning includes GAN encoders, fixed spherical targets, K-means pseudo-labels, and mutual-information objectives.These approaches focus primarily on representation learning rather than clustering itself.
- Self-Supervised Learning: Self-supervised learning constructs pretext tasks whose targets are computable without supervision and whose representations are intended to contain high-level semantic information.Examples include predicting patch locations and jigsaw permutations.
3. Deep Comprehensive Correlation Mining
DCCM combines pseudo-graph and pseudo-label supervision with transformation robustness and triplet mutual information to mine complementary correlations for unsupervised clustering. Highly confident information guides progressive representation learning, while triplet correlations add class-related feature correspondence.
- Framework overview: DCCM combines pseudo-graph, pseudo-label, local-robustness, and triplet mutual-information objectives in a unified unsupervised clustering framework.The framework jointly mines correlations among samples, category information, robustness to transformed inputs, and correspondence between deep and shallow features.
- Pseudo-graph supervision: The pseudo-graph uses cosine similarity between prediction features to label sufficiently similar sample pairs as same-class and dissimilar pairs as different-class.A fixed threshold constructs the graph, avoiding the noisy false-positive-pair disadvantage associated with dynamically decreasing thresholds.
- Pseudo-label supervision: Pseudo-label supervision extends pair-wise correlations by partitioning the pseudo-graph into K groups and assigning each sample its partition as a label.Only highly confident predictions, selected by a probability threshold, participate in pseudo-label training; the method targets one-hot predictions at the optimum.
- Local robustness: Feature-invariant loss aligns predictions for an original image and its geometrically transformed version, using transformed samples as easy positive pairs that stabilize training.Small perturbations make otherwise easy high-confidence samples more informative for parameter learning.
- Triplet mutual information: Triplet mutual information extends instance-level mutual information by maximizing correspondence between deep and shallow features of same-class samples while using positive and negative pairs selected from the pseudo-graph.This class-related feature-set formulation is intended to learn more discriminative features.
4. Experiments
Experiments evaluate DCCM across clustering, feature classification, ablations, and correlation analyses. DCCM outperforms comparison methods, while local robustness, pseudo-labels, and triplet mutual information each contribute to performance and progressively improving pseudo-graphs.
- Main Results: 0.623 clustering ACC on CIFAR-10 is 10.1% higher than DAC’s 0.522, while CIFAR-100 gains 8.9% over DAC.
- Main Results: DCCM achieves more discriminative CIFAR-10 embeddings than DAC, with separation improving across training epochs.
- Main Results: DCCM achieves better top-1 nonlinear classification results than VAE, AAE, BiGAN, NAT, and DIM on CIFAR-10 and CIFAR-100.
- Main Results: On CIFAR-10, DCCM’s convolutional and fully connected features exceed DIM by more than 8%.
- Correlation Analysis: Pseudo-labels improve clustering ACC by 7.1% over pseudo-graph supervision alone, while triplet mutual information adds a further 4.0%.
- Overall Study of DCCM: Pseudo-graph precision improves with training epochs, and nearly 80% of CIFAR-10 samples eventually have largest prediction probability above 0.9.
5. Conclusions
The conclusion presents DCCM as a deep unsupervised clustering framework that learns discriminative representations by mining comprehensive correlations. Experiments across challenging datasets and two tasks report significant improvement over state-of-the-art methods.
- DCCM mines sample correlations, feature mutual information, local robustness, and intercorrelations to learn discriminative representations.
- Experiments on several challenging datasets and two tasks report significant improvement over state-of-the-art methods.
6. Supplementary Material
Supplementary material provides proofs, metric definitions, architectures, additional comparisons, efficiency details, and pair-sampling analyses. These materials further document DCCM’s evaluation setting and implementation choices.
- Proofs: The supplementary proofs establish graph-partition behavior and the one-hot property of optimal softmax prediction features.
- Evaluation Metrics: Evaluation uses NMI, ARI, and ACC, with higher values indicating better clustering performance.
- Implementation: For CIFAR-10 and CIFAR-100, the architecture uses four convolutional layers, three pooling layers, and two fully connected layers.
- Additional Comparisons: Under the same architecture, DCCM achieves the best results on CIFAR-10 and CIFAR-100, while DeepCluster performs poorly with its released code.
- Efficiency: Training on CIFAR-100 takes 19 hours on a single GTX 1080Ti GPU, with faster hardware or multiple GPUs potentially reducing runtime.
- Pair Sampling: Pair-sampling strategies use n positive and n negative pairs per batch for nearly equal computational complexity, while costlier approaches yield negligible improvement.
- Pair Sampling: For triplet mutual information, positive pairs are noise-sensitive, harder negative pairs help training, and uniform sampling is important.