Learning Like Humans: Analogical Concept Learning for Generalized Category Discovery

Jizhou Han; Chenhao Ding; Yuhang He; Qiang Wang; Shaokun Wang; SongLin Dong; Yihong Gong

cs.CVMar 20, 2026

Learning Like Humans: Analogical Concept Learning for Generalized Category Discovery

Authors: Jizhou Han, Chenhao Ding, Yuhang He, Qiang Wang, Shaokun Wang, SongLin Dong, Yihong Gong

Organizations: State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University · School of Software Engineering, Xi’an Jiaotong University · Harbin Institute of Technology, Shenzhen · Shenzhen University of Advanced Technology

Abstract

Generalized Category Discovery (GCD) seeks to uncover novel categories in unlabeled data while preserving recognition of known categories, yet prevailing visual-only pipelines and the loose coupling between supervised learning and discovery often yield brittle boundaries on fine-grained, look-alike categories. We introduce the Analogical Textual Concept Generator (ATCG), a plug-and-play module that analogizes from labeled knowledge to new observations, forming textual concepts for unlabeled samples. Fusing these analogical textual concepts with visual features turns discovery into a visual-textual reasoning process, transferring prior knowledge to novel data and sharpening category separation. ATCG attaches to both parametric and clustering style GCD pipelines and requires no changes to their overall design. Across six benchmarks, ATCG consistently improves overall, known-class, and novel-class performance, with the largest gains on fine-grained data. Our code is available at: https://github.com/zhou-9527/AnaLogical-GCD.

Figures & tables

Panel of Figure 1 — Figure 1: Illustration of an example of human Analogical Learning mechanism and the proposed Analogical Textual Concept Generator (ATCG).

Figure 2: Overview of the AL-GCD Framework . The framework consists of two stages: (1) ATCG training, where the ATCG is trained using labeled images and text embeddings to acquire the ability to generate meaningful analogical text embeddings for unlabeled samples; (2) GCD Training, where ATCG generates text embeddings for unlabeled samples, which are fused with visual embeddings through a fusion head to produce fusion embeddings. These embeddings are optimized via contrastive learning.

Figure 3: Architecture of the Analogical Textual Concept Generator (ATCG), illustrating its input–output tensor structure.

(a) Clustering with the ground-truth number of categories $K$ given
	Method	CUB			Stanford Cars			FGVC Aircraft			CIFAR100			ImageNet100			Herbarium19
	Method	All	Old	New	All	Old	New	All	Old	New	All	Old	New	All	Old	New	All	Old	New
DINO	GCD (CVPR 22)	51.3	56.6	48.7	39.0	57.6	29.9	45.0	41.1	46.9	73.0	76.2	66.5	74.1	89.8	66.3	35.4	51.0	27.0
	GPC (ICCV 23)	55.4	58.2	53.1	42.8	59.2	32.8	46.3	42.5	47.9	77.9	85.0	63.0	76.9	94.3	71.0	-	-	-
	SimGCD (ICCV 23)	60.3	65.6	57.7	53.8	71.9	45.0	54.2	59.1	51.8	80.1	81.2	77.8	83.0	93.1	77.9	44.0	58.0	36.4
	CMS (CVPR 24)	68.2	76.5	64.0	56.9	76.1	47.6	56.0	63.4	52.3	82.3	85.7	75.5	84.7	95.6	79.2	36.4	54.9	26.4

Table 1: Comparison with the state of the art on GCD. The best results are in bold, the second best are underlined. † denotes reproduced results.

\triangle

denotes results from the CMS Appendix [ 6 ] .

(a) Clustering with the GT number of categories K given
Method	Fine-grained			Classification			All Datasets
	All	Old	New	All	Old	New	All	Old	New
CPT	65.9	71.3	63.2	85.3	88.3	83.7	68.9	75.3	65.6
GET	71.5	74.8	69.8	86.9	90.6	82.6	73.0	78.4	69.4
CMS △	64.7	74.5	59.9	83.1	89.5	76.3	66.1	76.5	59.6
+ AL-GCD	69.5	75.3	66.7	85.8	90.4	83.0	70.0	77.7	65.8

Table 2: Comparison on Fine-grained Avg, Classification Avg, and All Datasets Avg with CLIP backbone. The best values are in bold and the second best are underlined. † denotes reproduced results.

\triangle

denotes results from the CMS appendix [ 6 ] .

ATCG Initial Layer	ATCG Stacked Layers	CIFAR100			ImageNet100			CUB			Stanford Cars			FGVC Aircraft			Herbarium19
ATCG Initial Layer	ATCG Stacked Layers	All	Old	New	All	Old	New	All	Old	New	All	Old	New	All	Old	New	All	Old	New
✗	✗	81.1	84.2	75.2	89.9	94.2	87.5	69.6	75.8	66.5	69.4	82.2	62.9	53.5	57.9	51.7	47.9	62.3	39.9
✓	✗	83.7	84.3	82.7	90.4	94.9	88.0	73.4	76.7	71.8	74.9	86.1	69.0	56.1	56.5	55.9	48.8	62.9	41.3
✓	✓	84.7	84.5	84.9	92.6	95.3	91.2	74.7	79.6	72.3	78.3	86.3	74.4	58.6	61.5	57.2	50.3	63.6	43.1
Improv. over baseline		+3.6	+0.3	+9.7	+2.7	+1.1	+3.7	+5.1	+3.8	+5.8	+8.9	+4.1	+11.5	+5.1	+3.6	+5.5	+2.4	+1.3	+3.2

Table 3: Ablation study results for ATCG with various initial and stacked layer settings across different datasets.

Num. of ATCG Layers	CUB			CIFAR100
Num. of ATCG Layers	All	Old	New	All	Old	New
0	69.62	75.78	66.53	81.11	84.22	75.23
2	74.73	79.59	72.31	84.67	84.54	84.94
4	74.80	78.48	72.87	84.87	83.77	86.76
6	74.69	78.12	72.97	84.76	83.67	86.63

Table 4: Ablation study on the number of ATCG layers.

Figure 4: Ablation study of the $\alpha$ on Stanford Cars.

Generalized Category Discovery

Generalized category discovery learns known classes from labeled examples while grouping unlabeled data into known and previously unseen categories.

See all 26 papers

Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery
May 10, 2026Yulin Xu, Chunqi Guo, Yuanzhen Shuai +1Semi-Supervised Learning Generalized Category Discovery
University College London London, UK · Juniata College Huntingdon, USA
CoGe-GCD: Reframing Generalized Category Discovery with Compositional Generalization
Sep 9, 2026Luyao Tang, Jiewei Zheng, Kunze Huang +3Geometric Representation Learning Classification
The University of Hong Kong, Hong Kong SAR, China · Xiamen University, Xiamen, China · Shenzhen University, Shenzhen, China.
OmniGCD: Abstracting Generalized Category Discovery for Modality Agnosticism
Apr 16, 2026Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh +2Generalized Category Discovery
SAIVT, QUT, Australia · Shield AI, Australia · La Trobe University, Australia

May 10, 2026J/K move · Enter open · S save

Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery

Yulin Xu, Chunqi Guo, Yuanzhen Shuai, Jianyuan Ni

University College London London, UK · Juniata College Huntingdon, USA

Semi-Supervised Learning Generalized Category Discovery

Open paper AskPDF arXiv

In this study, we tackle Generalized Category Discovery (GCD) via a Relational Retrieval perspective, explicitly coupling labeled and unlabeled data through bidirectional knowledge transfer. While existing methods treat these sources separately, missing valuable interaction opportunities, we propose Relational Pattern Consistency (RPC) that enables mutual enhancement. RPC employs One-vs-All classifiers for soft ID/OOD decomposition, then introduces two mechanisms: (i) for known-class preservation, we transfer semantic behavioral alignment; (ii) for category discovery, we leverage the insight that samples from the same category maintain invariant relationships with known-class prototypes, transforming unreliable pseudo-labeling into well-defined relational pattern matching. This bidirectional design allows labeled data to guide unlabeled learning while discovering novel categories through their collective relational signatures. Extensive experiments demonstrate RPC achieves state-of-the-art performance on both generic and fine-grained benchmarks.

Abstract

Generalized Category Discovery

Explore similar work

Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery

CoGe-GCD: Reframing Generalized Category Discovery with Compositional Generalization

OmniGCD: Abstracting Generalized Category Discovery for Modality Agnosticism