Research map / Label-efficient and robust learning / Semi-supervised learning
RPSC: Robust Pseudo-Labeling for Semantic Clustering
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 RPSC, a two-stage deep clustering method that mines confidence-weighted semantic pseudo-labels with a soft contrastive loss (RPSC-Self) and then filters reliable labels for self-adaptive-threshold semi-supervised training (RPSC-Semi).
Problem
Deep clustering methods treat pseudo-labels as hard supervision regardless of confidence, so near-boundary samples with wrong semantic labels accumulate errors during training.
Approach
Assign each pseudo-label a confidence based on the sample's distance to its cluster center, use a soft contrastive loss that pulls samples toward high-confidence pseudo-labels and removes same-label pairs from negatives, and filter reliable labels via local neighbor consistency before semi-supervised training with FreeMatch thresholds.
Key results
- RPSC reaches 0.688 ACC on ImageNet-Dogs, up to 24% over the second-best method, per the abstract.
- RPSC-Semi on CIFAR-10 reaches NMI 0.875, ACC 0.940, ARI 0.876, improving over RPSC-Self (ACC 0.857).
- RPSC-Self alone outperforms prior baselines on most datasets but not on CIFAR-10 (SCAN ACC 0.876 vs RPSC-Self 0.857); RPSC-Semi fails on Tiny-ImageNet due to unreliable pseudo-labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RPSC instead of SPICE | image clustering · ImageNet-Dogs | ACC: 0.554 -> 0.688 | Helped |
| Use RPSC instead of SPICE | image clustering · CIFAR-10 | ACC: 0.926 -> 0.940 | Helped |
| Use RPSC instead of SPICE | image clustering · CIFAR-100-20 | ACC: 0.538 -> 0.556 | Helped |
| Use RPSC instead of SPICE | image clustering · STL-10 | ACC: 0.938 -> 0.950 | Helped |
| Use RPSC instead of SPICE | image clustering · ImageNet-10 | ACC: 0.959 -> 0.962 | Helped |
| Enable augmentation in the second branch while leaving the first branch unaugmented (ablation) | RPSC-Self · CIFAR-10 | NMI: 0.752 ± 0.003 -> 0.754 ± 0.003 | Helped |
| Enable augmentation in the first branch, with augmentation also in the second branch (ablation) | RPSC-Self · CIFAR-10 | NMI: 0.754 ± 0.003 -> 0.748 ± 0.005 | Hurt |
| Add the SCL denominator that excludes different pseudo-labels from the prediction (ablation) | SCL-1 · CIFAR-10 | NMI: 0.711 -> 0.732 | Helped |
| Include pseudo-label confidence in the SCL numerator (ablation) | SCL-3 · CIFAR-10 | NMI: 0.734 -> 0.754 | Helped |
| Replace cross-entropy with SCL | SCAN · CIFAR-100-20 | ACC: 0.459 -> 0.472 | Helped |
| Replace cross-entropy with SCL | SPICE-Self · CIFAR-100-20 | ACC: 0.468 -> 0.501 | Helped |
| Replace cross-entropy with SCL | RPSC-Self · CIFAR-100-20 | ACC: 0.486 -> 0.518 | Helped |
Surprising findings
- RPSC-Self does not surpass SCAN on CIFAR-10, which the authors attribute to backbone differences.
- RPSC-Semi cannot perform semi-supervised training on Tiny-ImageNet because RPSC-Self yields too few reliable pseudo-labels.
Limitations
- Performance on Tiny-ImageNet is poor because RPSC-Self produces too few reliable pseudo-labels for the semi-supervised stage.
- RPSC-Self underperforms SCAN on CIFAR-10.
Experimental setup
Datasets: STL-10 CIFAR-10 CIFAR-100-20 ImageNet-10 ImageNet-Dogs Tiny-ImageNet
Benchmarks: STL-10 CIFAR-10 CIFAR-100-20 ImageNet-10 ImageNet-Dogs Tiny-ImageNet
Metrics: NMI ACC ARI
Baselines: K-means SC AE DAE DCGAN DeCNN VAE JULE DEC DAC DDC IIC
Builds on: SPICE SCAN FreeMatch MoCo-v2 BYOL
Closest papers by meaning
- ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence (CVPR 2025 · similarity 0.88)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.87)
- Language-Assisted Image Clustering Guided by Discriminative Relational Signals and Adaptive Semantic Centers (NeurIPS 2026 · similarity 0.86)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (ICLR 2024 · similarity 0.86)
- Image Clustering with External Guidance (ICML 2024 · similarity 0.86)
- Improved Approximation Algorithms for Chromatic and Pseudometric-Weighted Correlation Clustering (NeurIPS 2025 · 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.