Research map / Label-efficient and robust learning / Recognition and long-tail
Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization
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 RLCD, which combines a reciprocal main/auxiliary-branch learning framework with class-wise distribution regularization to improve base discrimination and novel-class accuracy in parametric generalized category discovery.
Problem
Parametric generalized category discovery methods rely on unreliable self-supervision, leading to weak base-class discrimination and bias toward base classes that hurts novel-class accuracy.
Approach
An auxiliary AUX token trained only on pseudo-base samples provides base-class soft labels via KL distillation to the main all-class branch (reciprocal loop), and a class-wise distribution regularization loss matches batch-level class expectation matrices across two views to boost confidence and novel-class predictions.
Key results
- RLCD reaches 70.0 All accuracy on CUB200 with DINO ViT-B/16 versus 60.3 for SimGCD, a 3.6% average gain across fine-grained datasets on DINO.
- RLCD achieves 83.4 All accuracy on CIFAR100 and 86.9 on ImageNet-100 with DINO, the best reported in the table for those datasets, with average 89.2 vs 86.7 for SimGCD.
- With DINOv2, RLCD averages 72.8 across the four fine-grained datasets, vs 67.2 for SimGCD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RLCD instead of the strongest listed baseline | GCD with DINO ViT-B/16 · CUB200 | All accuracy: 69.4 -> 70.0 | Helped |
| use RLCD instead of the strongest listed baseline | GCD with DINO ViT-B/16 · Stanford Cars | All accuracy: 59.0 -> 64.9 | Helped |
| use RLCD instead of the strongest listed baseline | GCD with DINO ViT-B/16 · CIFAR10 | All accuracy: 97.7 -> 97.4 | Hurt |
| add CDR to the Main-only loss (ablation) | RLCD training on CUB200 · CUB200 | Novel accuracy: 57.7 -> 63.0 | Helped |
| add CDR to the Main-only loss (ablation) | RLCD training on CUB200 · CUB200 | OB accuracy: 83.9 -> 83.7 | No clear effect |
| add the AUX branch (ablation) | Main-loss GCD on CUB200 · CUB200 | Base accuracy: 70.8 -> 73.7 | Helped |
| add cross-branch distillation (ablation) | Main + AUX on CUB200 · CUB200 | Base accuracy: 73.7 -> 76.4 | Helped |
| apply CDR to the main branch (ablation) | Main + AUX + Distill on CUB200 · CUB200 | Novel accuracy: 61.7 -> 65.2 | Helped |
| apply CDR to both branches rather than only the main branch (ablation) | RLCD on CUB200 · CUB200 | Base accuracy: 78.1 -> 79.1 | Helped |
| remove cross-branch distillation (ablation) | RLCD on CUB200 · CUB200 | Base accuracy: 79.1 -> 71.7 | Hurt |
| remove the AUX branch (ablation) | RLCD on CUB200 · CUB200 | Base accuracy: 79.1 -> 73.8 | Hurt |
| use maximum auxiliary probability instead of zero uncertainty weight (ablation) | cross-branch distillation on CUB200 · CUB200 | Base accuracy: 71.7 -> 79.1 | Helped |
Surprising findings
- The InfoNCE loss in SimGCD pushes apart same-class features and was removed in RLCD, which the authors argue improves discrimination.
- Standard regularizers such as entropy minimization severely hurt GCD performance, while the proposed CDR improves novel-class accuracy.
Limitations
- Gains on CIFAR10 are on par with existing methods rather than clearly better.
- Performance drops when the number of categories is estimated (e.g., 2.5% on Stanford Cars with GCD's estimate), so the method still depends on category count.
- Only the last ViT block and AUX token are trained; evaluated only on image classification with ViT backbones.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet-100 CUB200 Stanford Cars FGVC-Aircraft Herbarium-19
Benchmarks: CIFAR10 CIFAR100 ImageNet-100 CUB200 Stanford Cars FGVC-Aircraft Herbarium-19
Metrics: clustering accuracy (ACC) All/Base/Novel oracle base accuracy (OB)
Baselines: k-means RS+ UNO+ ORCA CRNCD GCD DCCL GPC SimGCD uGCD CMS InfoSieve
Builds on: SimGCD SupCon GCD CRNCD SPTNet
Closest papers by meaning
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025 · similarity 0.91)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.90)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.90)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.90)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.89)
- Towards Understanding Parametric Generalized Category Discovery on Graphs (ICML 2025 · similarity 0.89)
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.