Research map / Label-efficient and robust learning / Recognition and long-tail
Joint Class-level and Instance-level Relationship Modeling for Novel Class Discovery
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 symmetric class-to-class relationship modeling via logit-based transition matrices with cyclic transition, plus decoupled pairwise-similarity consistency constraints, for novel class discovery.
Problem
Novel class discovery must cluster unlabeled samples from unknown classes using a labeled set of related but different known classes; existing methods model knowledge transfer only via shared representations or coarse-grained/asymmetric relationships, neglecting symmetric class-level relations.
Approach
Builds directional transition matrices between known and unknown classes from predicted logits, composes them into cyclic transitions, and maximizes self-transition probabilities for bidirectional class-level transfer; adds KL-based consistency between labeled-logit and unlabeled-logit pairwise similarities for instance-level transfer.
Key results
- On CIFAR100-50 task-aware clustering, the method reaches 68.2% accuracy, outperforming UNO (62.3%) by 5.9%, IIC by 2.4%, and CrKD by 2.9%.
- On CUB task-aware clustering, the method reaches 69.1%, outperforming UNO (59.2%) by 9.9 points.
- Ablation on CIFAR100-50 shows the class-level and instance-level modules each improve over UNO (ACC 62.3% to 67.0% and 63.1%), with the combination achieving 68.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add class-level relationship modeling (ablation) | Ours · CIFAR100-50 unlabeled training set | ACC: 62.3 -> 67.0 | Helped |
| add instance-level relationship modeling (ablation) | Ours · CIFAR100-50 unlabeled training set | ACC: 62.3 -> 63.1 | Helped |
| add instance-level relationship modeling to class-level modeling (ablation) | Ours · CIFAR100-50 unlabeled training set | ACC: 67.0 -> 68.2 | Helped |
| add class-level relationship modeling to instance-level modeling (ablation) | Ours · CIFAR100-50 unlabeled training set | ACC: 63.1 -> 68.2 | Helped |
| use joint class-level and instance-level relationship modeling instead of CrKD | novel class discovery · CIFAR100-50 unlabeled test set (task-agnostic) | clustering accuracy (%): 59.4 -> 64.2 | Helped |
| use joint class-level and instance-level relationship modeling instead of CrKD | novel class discovery · CIFAR100-50 labeled test set (task-agnostic) | classification accuracy (%): 78.6 -> 76.3 | Hurt |
| use joint class-level and instance-level relationship modeling instead of CrKD | novel class discovery · CUB unlabeled training set (task-aware) | clustering accuracy (%): 65.7±0.6 -> 69.1±2.7 | Helped |
| use joint class-level and instance-level relationship modeling instead of CrKD | novel class discovery · Stanford Cars unlabeled training set (task-aware) | clustering accuracy (%): 53.5±0.8 -> 53.5±2.0 | No clear effect |
Surprising findings
- On Stanford Cars and FGVC-Aircraft the gains over IIC and CrKD are small, which the authors attribute to generally high semantic similarity among classes.
Limitations
- Some reported baselines (IIC, CrKD on some splits) were reproduced by the authors, introducing comparison uncertainty.
- Gains over IIC and CrKD on Stanford Cars and Aircraft are marginal.
- Assumes the number of unknown classes is known.
Experimental setup
Datasets: CIFAR100 Stanford Cars CUB FGVC-Aircraft ImageNet Herbarium-19
Benchmarks: CIFAR100-50 CIFAR100-80 Stanford Cars CUB FGVC-Aircraft ImageNet-1k ImageNet-100 Herbarium-19
Metrics: clustering accuracy classification accuracy NMI ARI
Baselines: K-means DTC RS+ NCL UNO ComEx IIC CrKD SNCD
Builds on: UNO CrKD IIC RS+
Closest papers by meaning
- A Unified Knowledge Transfer Network for Generalized Category Discovery (AAAI 2024 · similarity 0.88)
- Semantic-Guided Novel Category Discovery (AAAI 2024 · similarity 0.86)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.86)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.85)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.85)
- Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation (CVPR 2024 · similarity 0.84)
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.