Research map / Label-efficient and robust learning / Recognition and long-tail
Hyperbolic Category 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 HypCD, which maps the Euclidean features of self-supervised backbones (DINO/DINOv2) into the Poincaré ball and trains hyperbolic representation and classifier heads for generalized category discovery, improving GCD, SimGCD and SelEx on SSB and generic benchmarks.
Problem
Generalized category discovery on partially labelled image datasets; existing GCD methods embed features in Euclidean or spherical space, which poorly capture hierarchical structure of object parts and categories, limiting knowledge transfer from known to unknown classes.
Approach
Map backbone features into hyperbolic (Poincaré ball) space via feature clipping and exponential mapping, then learn representations with a hybrid distance- and angle-based contrastive loss and a hyperbolic linear classifier (HypFFN), plugged into existing GCD baselines.
Key results
- Hyp-SelEx reaches 79.8% All ACC on CUB with DINO (vs. 73.6% for SelEx) and 90.7% with DINOv2.
- Hyp-GCD improves Stanford-Cars All ACC by 11.8 points over the GCD baseline with DINO.
- Hyp-SimGCD reaches 91.5% All ACC on CIFAR-100 with DINOv2, and hyperbolic variants consistently narrow the Old-New accuracy gap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply HypCD to SelEx | SelEx with DINO · CUB | ACC: 73.6 -> 79.8 | Helped |
| Apply HypCD to SelEx | SelEx with DINO · Stanford-Cars | ACC: 58.5 -> 62.9 | Helped |
| Apply HypCD to SelEx | SelEx with DINO · FGVC-Aircraft | ACC: 57.1 -> 65.9 | Helped |
| Apply HypCD to SelEx | SelEx with DINO · CIFAR-10 | ACC: 95.9 -> 96.7 | Helped |
| Apply HypCD to SelEx | SelEx with DINO · CIFAR-100 | ACC: 82.3 -> 82.4 | Helped |
| Apply HypCD to SelEx | SelEx with DINO · ImageNet-100 | ACC: 83.1 -> 86.8 | Helped |
| Apply HypCD to SelEx | SelEx with DINOv2 · CUB | ACC: 87.4 -> 90.7 | Helped |
| Apply HypCD to SelEx | SelEx with DINOv2 · Stanford-Cars | ACC: 82.2 -> 83.8 | Helped |
| Apply HypCD to SelEx | SelEx with DINOv2 · FGVC-Aircraft | ACC: 79.8 -> 83.4 | Helped |
| Apply HypCD to SelEx | SelEx with DINOv2 · CIFAR-10 | ACC: 98.5* -> 98.6 | Helped |
| Apply HypCD to SelEx | SelEx with DINOv2 · CIFAR-100 | ACC: 87.7* -> 88.6 | Helped |
| Apply HypCD to SelEx | SelEx with DINOv2 · ImageNet-100 | ACC: 90.9* -> 92.3 | Helped |
Surprising findings
- Hyperbolic-trained features used with Euclidean k-means for non-parametric methods perform comparably to hyperbolic-space clustering while being more efficient.
Limitations
- Optimal curvature, clipping value and loss weight differ between generic and fine-grained datasets, requiring dataset-specific tuning.
- Gains on some datasets (e.g., CIFAR-10) are marginal because performance is near saturation.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-100 CUB Stanford-Cars FGVC-Aircraft
Benchmarks: SSB (CUB, Stanford-Cars, FGVC-Aircraft) CIFAR-10 CIFAR-100 ImageNet-100
Metrics: ACC (All, Old, New)
Baselines: ORCA GCD XCon OpenCon PromptCAL DCCL GPC CiPR SimGCD µGCD InfoSieve SPTNet
Builds on: GCD SimGCD SelEx hyperbolic linear layer (Chen et al. 2021) DINO
Closest papers by meaning
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.93)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.89)
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025 · similarity 0.88)
- SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery (NeurIPS 2025 · similarity 0.87)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.87)
- Towards Understanding Parametric Generalized Category Discovery on Graphs (ICML 2025 · similarity 0.87)
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.