Research map / Label-efficient and robust learning / Self-supervised and contrastive
Clustering via Self-Supervised Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CLUDI, a self-supervised deep clustering method that uses a conditional diffusion model over assignment embeddings on frozen DINO ViT features, trained with a teacher-student (SimSiam-style) setup and a uniform-prior non-collapsing cross-entropy loss.
Problem
Deep clustering / unsupervised classification of image features often suffers from model collapse and underuses pre-trained features; diffusion models had not been applied to clustering.
Approach
Use a data-conditioned diffusion model to generate stochastic assignment embeddings from noise, mapped to cluster probabilities via a linear head and tempered softmax; the teacher's diffusion sampling randomness acts as a data augmentation for a student trained with embedding MSE plus a non-collapsing cross-entropy over mini-batches.
Key results
- On ImageNet 50, CLUDI (ViT-B/16) reports 91.2% NMI, 82.1% ACC, and 76.2% ARI, versus 86.1/80.1/71.0 for TEMI (ViT-B/16).
- On CIFAR-10 with ViT-B/16, CLUDI reports 89.6% NMI, 95.3% ACC, 89.8% ARI, slightly above TEMI (88.6/94.5/88.5).
- On Oxford 102 Flower with ViT-B/16, CLUDI reports 69.7% ACC versus 25.9% for TEMI (ViT-B/16) and 57.8% for Self-Classifier* (ViT-B/16).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLUDI instead of the strongest listed baseline | ViT-B/16 clustering · CIFAR-10 | NMI (%): 88.6 -> 89.6 | Helped |
| Use CLUDI instead of the strongest listed baseline | ViT-B/16 clustering · STL-10 | ACC (%): 98.5 -> 98.7 | Helped |
| Use CLUDI instead of the strongest listed baseline | Oxford-IIIT Pets clustering · Oxford-IIIT Pets | ARI (%): 63.0 -> 71.6 | Helped |
| Use CLUDI instead of the strongest listed baseline | ViT-B/16 clustering · Oxford 102 Flower | ACC (%): 57.8 -> 69.7 | Helped |
| Use CLUDI instead of the strongest listed baseline | ViT-B/16 clustering · Caltech 101 | NMI (%): 83.5 -> 87.9 | Helped |
| Use CLUDI instead of the strongest listed baseline | ViT-B/16 clustering · ImageNet 50 | NMI (%): 87.8 -> 91.2 | Helped |
| Use CLUDI instead of the strongest listed baseline | ViT-B/16 clustering · ImageNet 100 | NMI (%): 85.6 -> 87.1 | Helped |
| Vary the classification-loss weight λ (ablation) | CLUDI · ImageNet 100 validation | NMI, ACC, ARI | Mixed |
| Vary embedding dimension d (ablation) | CLUDI · ImageNet 100 validation | NMI, ACC, ARI | Mixed |
| Vary diffusion noise rescaling factor F² (ablation) | CLUDI · ImageNet 100 validation | NMI, ACC, ARI | Mixed |
Limitations
- Performance depends on sensitive hyperparameters (noise rescaling F^2, embedding dimension d, loss weight lambda) that were tuned on the tested datasets.
- Performance may degrade when scaling to a large number of clusters K.
- Relies on frozen pre-trained DINO features; baselines were not all re-run under identical conditions (some results taken from original papers).
Experimental setup
Datasets: ImageNet 50 ImageNet 100 ImageNet 200 Oxford-IIIT Pets Oxford 102 Flower Caltech 101 CIFAR-10 STL-10
Benchmarks: CIFAR-10 STL-10 ImageNet 50 ImageNet 100 ImageNet 200 Oxford-IIIT Pets Oxford 102 Flower Caltech 101
Metrics: NMI ACC ARI
Baselines: SCAN Propos TSP TEMI Self-Classifier
Builds on: DDIM SimSiam DINO TEMI Min-SNR-gamma
Closest papers by meaning
- Sparsely Supervised Diffusion (NeurIPS 2026 · similarity 0.86)
- SODA: Bottleneck Diffusion Models for Representation Learning (CVPR 2024 · similarity 0.86)
- Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.85)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.85)
- Breaking the Reclustering Barrier in Centroid-based Deep Clustering (ICLR 2025 · similarity 0.85)
- Improving Training Efficiency of Diffusion Models via Multi-Stage Framework and Tailored Multi-Decoder Architecture (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.