Dynamic Conceptional Contrastive Learning for Generalized Category Discovery

• GCD requires distinguishing known from unknown classes while discovering novel clusters without annotations.Instance-only self-supervision overlooks relationships among samples sharing class-, super-class-, or sub-class-level conceptions, degrading representation learning.Source: §1 · p. 2
• DCCL alternates Dynamic Conception Generation and Dual-level Contrastive Learning to leverage relationships between unlabeled data for representation learning.It dynamically generates conceptions with semi-supervised conceptional consolidation, then optimizes instance- and conception-level objectives while maintaining a dynamic memory of up-to-date conceptions.Source: §1 · p. 2
• A dynamic conception generation and update mechanism is designed to ensure consistent conception learning and encourage more discriminative representations.Source: §1 · p. 2
• DCCL consistently outperforms state-of-the-art GCD algorithms on both generic and fine-grained tasks.Source: §1 · p. 2

2.1. Novel Category Discovery

• Prior NCD and GCD methods often assume the clustered data’s class count is known, while DTC and GCD search for it independently of representation learning.DCCL jointly considers downstream clustering and representation learning; experiments show the two are mutually beneficial.Source: §2.1 · p. 2

2.2. Contrastive Learning based on Memory Buffer

• Unlike PCL and SCL, which use fixed class counts during training, DCCL dynamically estimates the number of classes at different training stages.PCL also requires an instance-level memory buffer to produce prototypes, while SCL considers only three fixed class counts.Source: §2.2 · p. 2

2.3. Semi-Supervised Learning and Clustering

• Unlike SSL’s same-class assumption, semi-supervised clustering can use varied supervision, but prior methods described here do not adaptively assign conception counts.DCCL improves InfoMap to use labeled and unlabeled data and produce varying conceptional labels across contrastive-learning epochs.Source: §2.3 · p. 3

3.1. Problem Formulation

• GCD clusters partially labeled data whose unlabeled images may come from either known or unknown classes.The number of unlabeled classes is unknown during training.Source: §3.1 · p. 3

3.2. Overview

• DCCL alternates DCG and DCL, with each stage providing guidance that improves the other for downstream clustering.DCG generates conceptions while DCL learns representations using supervised information; DCG then provides increasingly comprehensive guidance as conceptual relationships are learned.Source: §3.2 · pp. 3–4

3.3. Dynamic Conception Generation

• DCG uses labeled-data similarity constraints to build a network, then applies hyper-parameter-free Infomap clustering to assign conceptions.High-confidence links are selected by a threshold, and similarities between labeled positive pairs use their maximum neighborhood similarity rather than 1 to avoid grouping all such instances into one cluster.Source: §3.3 · p. 4
• The conceptional memory buffer stores conception representations initialized as the mean feature of instances sharing each conception label.The number of conceptions changes during training, and momentum updates help maintain conceptional consistency.Source: §3.3 · p. 4, §3.4 · p. 5

3.4. Dual-Level Contrastive Learning

• DCL pulls each instance representation toward its assigned conception and away from other conceptions, while dispersion loss pushes different conceptions apart.The dispersion loss filters high-uncertainty conception pairs; instance-level contrastive learning combines supervised and self-supervised losses.Source: §3.4 · p. 5

3.5. Joint Optimization

• Training repeatedly performs DCG and DCL through the maximum epoch and optimizes a weighted sum of loss functions.The loss weights adjust the strengths of the two loss functions.Source: §3.5 · p. 5

4.1. Experimental Setup

• DCCL is evaluated on three generic and three fine-grained datasets, with half the classes designated “Old” and half the instances of each labeled class included in the labeled set.The datasets are CIFAR-10, CIFAR-100, ImageNet-100, CUB-200, Stanford Cars, and Oxford-IIIT Pet.Source: §4.1 · p. 5, Table 1 · p. 6

4.3. Effectiveness Evaluation

• Adding dispersion loss improves CUB-200 “New” performance by 2.7%; momentum updating improves old and new classes, while conception consolidation rectifies estimated seen–unseen relations.Conception-level contrastive learning alone performs competitively against the instance-level baseline.Source: Table 4 · p. 7, Figure 4 · p. 7, §4.3 · p. 7

4.4. Hyper-Parameter Analyses

• The selected settings are α = 0.3, β = 0.1, τ_M = 0.3, τ_I = 5, and η = 0.9; smoother CMB running averages perform better.Loss weights are selected using held-off validation data.Source: §4.4 · pp. 7–8

4.5. Further Investigation of DCG

• DCG starts with fewer coarse conceptions and later refines relations; its conception count nears class count without targeting it, while known-k SSK produces worse clustering.Generic datasets tend to yield more conceptions than classes, while fine-grained datasets tend to yield fewer; the authors suggest subclass and superclass conceptions may help, respectively.Source: §4.5 · p. 8, Table 5 · p. 8

4.6. Evaluation with Different Split Protocols

• On varying CIFAR-100 splits, DCCL is more robust when few labeled classes are available, while GCD degrades as unlabeled classes increase.The evaluation varies the splits to test performance under strict annotation limitations.Source: §4.6 · p. 8

• DCCL achieves new state-of-the-art performance on GCD tasks.Source: §5 · p. 8
• DCCL alternates between exploring latent conceptional relationships and performing conceptional contrastive learning to learn more discriminable representations.Source: §5 · p. 8
• A conception-level momentum update addresses inconsistent conception-representation updates during training and facilitates optimization.Source: §5 · p. 8