Research map / Label-efficient and robust learning / Semi-supervised learning
Unsupervised Learning for Class Distribution Mismatch
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 UCDM, an unsupervised class-distribution-mismatch method that uses Stable Diffusion 2.0 to synthesize positive (add class) and negative (erase class) image pairs from unlabeled seeds, plus confidence-based pseudo-labeling of real images.
Problem
Class distribution mismatch in training data (unknown and new classes) makes semi-supervised open-set classifiers rely on labeled data; the paper asks how to handle CDM without any ground-truth labels, using only known class names.
Approach
Use a text-to-image diffusion model with conditional DDIM inversion to erase a known class from a seed image (negative) and class-prompted generation to add it (positive), building unsupervised positive-negative pairs for a one-vs-all open-set classifier, then iteratively pseudo-label confident real images.
Key results
- On Tiny-ImageNet at 60% mismatch, UCDM without labels surpasses OpenMatch (40 labels per class) by 5.0%, 91.4%, and 89.5% on known, unknown, and new class accuracy (per the intro; Table 4 values differ).
- UCDM attains the highest balance score in all open-set settings on CIFAR-10, CIFAR-100 and Tiny-ImageNet across 20%-75% mismatch (e.g., 94.6 on CIFAR-10 at 60%).
- Ablation on CIFAR-10 at 60% mismatch: full UCDM gets 95.6 known-class accuracy and 94.6 balance, versus 84.8 and 71.5 with random-noise instances instead of the positive/negative pipeline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UCDM instead of OpenMatch | Tiny-ImageNet, 60% mismatch open-set task · Tiny-ImageNet, 60% mismatch | unknown-class accuracy: 3.5 -> 94.9 | Helped |
| Use UCDM instead of OpenMatch | Tiny-ImageNet, 60% mismatch open-set task · Tiny-ImageNet, 60% mismatch | new-class accuracy: 5.9 -> 95.4 | Helped |
| Use UCDM instead of OpenMatch | Tiny-ImageNet, 60% mismatch open-set task · Tiny-ImageNet, 60% mismatch | known-class accuracy: 10.8 -> 15.8 | Helped |
| Use UCDM instead of IOMatch | CIFAR-100, 60% mismatch open-set task · CIFAR-100, 60% mismatch | balance score: 9.0 -> 47.2 | Helped |
| Use UCDM instead of CCSSL | CIFAR-10, 60% mismatch closed-set task · CIFAR-10, 60% mismatch | accuracy: 95.7 -> 95.6 | No clear effect |
| Add the positive-generation pipeline (ablation) | UCDM · CIFAR-10, 60% mismatch | balance score: 72.8 -> 94.6 | Helped |
| Add the negative-generation pipeline (ablation) | UCDM · CIFAR-10, 60% mismatch | balance score: 79.3 -> 94.6 | Helped |
| Use the proposed positive and negative generation pipelines rather than random noise (ablation) | UCDM · CIFAR-10, 60% mismatch | balance score: 71.5 -> 94.6 | Helped |
| Add the real-image loss term for known pseudo-labeled data (ablation) | UCDM learning objectives · CIFAR-10, 60% mismatch | known-class accuracy | Helped |
| Use a confidence threshold below 0.95 (ablation) | UCDM confidence-based labeling · CIFAR-10 | open-set performance | Hurt |
Surprising findings
- On the closed-set task UCDM is not the top performer at 0% mismatch; the authors attribute this to fewer generated positives when the training set is smallest.
- SCDM-style baselines like OpenMatch and IOMatch show unstable open-set performance, with OpenMatch's balance score dropping sharply on several settings.
Limitations
- Prompt variability is limited, which restricts diversity of positive instances.
- Performance is sensitive to the confidence threshold; below 0.95 incorrect pseudo-labels degrade open-set results.
- Experiments are limited to small image benchmarks (CIFAR and Tiny-ImageNet) with WideResNet-28-2.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet
Metrics: known-class accuracy unknown-class accuracy new-class accuracy balance score
Baselines: UASD DS3L T2T CCSSL MTCF OpenMatch IOMatch
Builds on: DDIM Stable Diffusion OpenMatch FixMatch-style SSL
Closest papers by meaning
- CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.89)
- Improving Barely Supervised Learning by Discriminating Unlabeled Samples with Super-Class (NeurIPS 2022 · similarity 0.87)
- Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 2024 · similarity 0.85)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.85)
- Out-Of-Domain Unlabeled Data Improves Generalization (ICLR 2024 · similarity 0.85)
- Confusing Pair Correction Based on Category Prototype for Domain Adaptation under Noisy Environments (AAAI 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.