Research map / Label-efficient and robust learning / Semi-supervised learning
Targeted Representation Alignment for Open-World Semi-Supervised Learning
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 TRAILER, which aligns features to a pre-assigned simplex ETF (neural collapse) with progressive prototypical calibration and a coarse-to-fine sample-target allocation (optimal transport plus PU-learning label refinery) for open-world semi-supervised learning.
Problem
Open-world semi-supervised learning classifies unlabeled data drawn from both known and novel classes; similarity-based clustering suffers from representation collapse where novel class features become inseparable.
Approach
Use a fixed targeted classifier pre-assigned to a simplex equiangular tight frame to induce neural-collapse-style feature arrangement, progressively blended with prototypical clustering, and refine unreliable pseudo-labels via optimal transport and a PU-learning known/novel separation head.
Key results
- On CIFAR-10, TRAILER reaches 94.4% overall accuracy, improving the best baseline by 3.1%.
- On CIFAR-100 and ImageNet-100, TRAILER gets 55.6% and 85.3% overall accuracy, gains of 2.9% and 1.5% over the best baselines.
- On fine-grained SSB benchmarks, TRAILER improves novel-class accuracy by 8.8%, 14.8%, 2.6%, and 9.2% on CUB, Stanford Cars, FGVC-Aircraft, and Herbarium 19.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRAILER instead of NACH | open-world SSL · CIFAR-10 | All accuracy: 91.3 -> 94.4 | Helped |
| Use TRAILER instead of OpenCon | open-world SSL · CIFAR-100 | All accuracy: 52.7 -> 55.6 | Helped |
| Use TRAILER instead of OpenCon | open-world SSL · ImageNet-100 | All accuracy: 83.8 -> 85.3 | Helped |
| Use TRAILER instead of GPC | open-world SSL · CUB | All accuracy: 55.4 -> 65.1 | Helped |
| Use TRAILER instead of OpenCon | open-world SSL · Stanford Cars | All accuracy: 49.1 -> 55.4 | Helped |
| Use TRAILER instead of GPC | open-world SSL · FGVC-Aircraft | All accuracy: 46.3 -> 54.5 | Helped |
| Use TRAILER instead of GPC | open-world SSL · Herbarium 19 | All accuracy: 36.5 -> 44.5 | Helped |
| Add representation calibration with a targeted classifier (ablation) | TRAILER · CIFAR-10 | All accuracy: 70.6 -> 94.4 | Helped |
| Add representation calibration with a targeted classifier (ablation) | TRAILER · CIFAR-100 | All accuracy: 33.4 -> 55.6 | Helped |
| Use progressive calibration instead of setting λ = 0 (ablation) | TRAILER · CIFAR-10 | All accuracy: 92.5 -> 94.4 | Helped |
| Use progressive calibration instead of setting λ = 0 (ablation) | TRAILER · CIFAR-100 | All accuracy: 54.6 -> 55.6 | Helped |
| Add optimal-transport rough assignment instead of vanilla softmax assignment (ablation) | TRAILER · CIFAR-10 | All accuracy: 92.5 -> 94.4 | Helped |
Surprising findings
- Removing representation calibration causes novel-class accuracy on CIFAR-100 to collapse from 48.7% to 14.9%, with known accuracy barely changing.
Limitations
- Relies on the number of novel classes being known a priori; estimated K gives lower performance on CIFAR-100 (53.3 vs 55.6 overall).
- Pseudo-label quality in early training is noisy, requiring a confidence-based filtering schedule and warm-up.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-100 CUB-200 Stanford Cars FGVC-Aircraft Herbarium 19
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-100 CUB-200 Stanford Cars FGVC-Aircraft Herbarium 19
Metrics: known class accuracy novel class accuracy all-class accuracy (Hungarian matched)
Baselines: FixMatch DS3L CGDL DTC RankStats ORCA GCD OpenLDN OpenCon NACH OpenNCD UNO+
Builds on: ORCA OpenCon UNO SimCLR Sinkhorn-Knopp
Closest papers by meaning
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.87)
- OwMatch: Conditional Self-Labeling with Consistency for Open-World Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning (ICML 2024 · similarity 0.86)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.86)
- Learning Textual Prompts for Open-World Semi-Supervised Learning (CVPR 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.