Dynamic Conceptual Contrastive Learning (DCCL)

• DCCL is a framework that dynamically estimates latent concept structures during training to supervise contrastive representation learning.
• It alternates between dynamic conception generation and dual-level contrastive objectives, effectively handling partially labeled datasets with both known and unknown categories.
• Empirical results show that DCCL improves clustering accuracy and performance on benchmarks like CIFAR and CUB-200 by adaptively refining fine-grained conceptual groupings.

Dynamic Conceptual Contrastive Learning (DCCL) denotes a class of contrastive-learning procedures in which latent concept structure is estimated during training and then used to supervise representation learning. In the generalized category discovery literature, the formulation originates in "Dynamic Conceptional Contrastive Learning for Generalized Category Discovery" (Pu et al., 2023), where the method alternates between Dynamic Conception Generation (DCG) and Dual-level Contrastive Learning (DCL) to model latent visual groupings at class, super-class, and sub-class granularity. A later line of work, most notably the MGCE framework, uses the spelling "Dynamic Conceptual Contrastive Learning" for a closely related module that alternates concept mining and dual-level representation learning while allowing the number of unlabeled categories to remain unknown during training (Zheng et al., 30 Sep 2025). Across these variants, the defining principle is that concept assignments are not fixed annotations but evolving latent variables coupled to the representation learner.

1. Terminology and scope

The acronym DCCL is not unique in the literature. In the context most directly associated with dynamic conceptual learning, it refers either to "Dynamic Conceptional Contrastive Learning" for generalized category discovery (Pu et al., 2023) or to the DCCL module inside the later "Multi-Granularity Conceptual Experts" framework (Zheng et al., 30 Sep 2025). Other papers reuse the same acronym for different methods and problem settings.

arXiv id	Expansion of DCCL	Domain
(Pu et al., 2023)	Dynamic Conceptional Contrastive Learning	Generalized Category Discovery
(Zheng et al., 30 Sep 2025)	Dynamic Conceptual Contrastive Learning	Generalized Fine-Grained Category Discovery
(Zhao et al., 2023)	Disentangled Causal Embedding with Contrastive Learning	Recommender Systems
(Wei et al., 19 Oct 2025)	Domain-Connecting Contrastive Learning	Domain Generalization
(Long et al., 2022)	Domain Confused Contrastive Learning	Unsupervised Domain Adaptation

This terminological overlap is a recurring source of confusion. In the generalized category discovery sense, DCCL is a non-parametric or semi-parametric framework for discovering latent concepts and using them as supervision. In the recommender, domain generalization, and domain adaptation usages, the acronym refers to distinct causal or domain-bridging objectives rather than concept mining. A careful reading of the title and task setting is therefore necessary before attributing results or mechanisms to "DCCL."

2. Problem setting in generalized category discovery

The original DCCL formulation is built for Generalized Category Discovery (GCD), a partially labeled setting in which unlabeled data contain both known and novel categories (Pu et al., 2023). Using the notation of the paper, the dataset is

$\mathcal{D} = \mathcal{D}^L \cup \mathcal{D}^U,$

with labeled data

$$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$

and unlabeled data

$$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$

where only a subset of classes are labeled and $\mathcal{Y}^L \subset \mathcal{Y}^U$. The number of labeled classes $N^L = |\mathcal{Y}^L|$ is known, whereas the number of unlabeled classes $N^U = |\mathcal{Y}^U|$ is unknown during training (Pu et al., 2023).

The later MGCE formulation expresses the same problem with labeled and unlabeled sets

$$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$

label spaces $Y_L \subset Y_U$, and an unlabeled class count $K_U$ that may be unknown during training (Zheng et al., 30 Sep 2025). This continuity of notation reflects a stable problem definition: the learner must preserve performance on known categories while discovering the latent structure of novel ones.

The evaluation protocol in the original DCCL work uses clustering accuracy (ACC) after label assignment, reported for All, Old, and New categories (Pu et al., 2023). This reporting convention is important because a method can improve discovery of unseen classes while degrading discrimination among seen ones, or vice versa. The DCCL design explicitly targets that trade-off by injecting labeled constraints into concept estimation while still learning from the full partially labeled dataset.

3. Dynamic concept generation

A defining feature of DCCL is that concept labels are not treated as static pseudo-labels. They are re-estimated from evolving features by a Dynamic Conception Generation step (Pu et al., 2023). The original method first computes cosine-based similarities

$$s_{ij} = \frac{\left(\frac{v_i}{\|v_i\|}\right)\cdot \left(\frac{v_j}{\|v_j\|}\right) + 1}{2} \in [0,1], \qquad s_i^{\max} = \max_{j \in \mathcal{D}} s_{ij},$$

then builds an adjacency matrix $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$0 with conception consolidation: $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$1 The use of $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$2 rather than $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$3 for same-label edges is explicitly motivated as a way to avoid collapsing all labeled positives into a single oversized cluster in Infomap (Pu et al., 2023).

Infomap is then run on this graph to produce hard concept assignments

$$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$4

and a concept-labeled feature set

$$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$5

If the current concept vocabulary has size $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$6, concept prototypes are initialized as

$$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$7

The entire DCG procedure is re-run every $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$8 epochs on center-cropped images (Pu et al., 2023).

The later MGCE version retains the same alternating logic but strengthens graph construction through Semi-Infomap, a KNN graph, and an adaptive search over $$\mathcal{D}^L = \{ (x_i^L, y_i^L) \}_{i=1}^{M^L} \subset \mathcal{X} \times \mathcal{Y}^L$$9 (Zheng et al., 30 Sep 2025). Its similarity is again cosine-based,

$$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$0

and the graph keeps only KNN neighbors after thresholding. The adaptive selection uses a coarse search over

$$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$1

followed by a fine search scored by

$$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$2

When $$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$3 is unknown, clusters with fewer than $$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$4 samples are treated as noise during inference (Zheng et al., 30 Sep 2025).

The original DCCL paper reports a characteristic training trajectory: early epochs produce fewer conceptions, which are described as coarser and super-class-like, whereas later epochs produce more fine-grained conceptions as features improve (Pu et al., 2023). This dynamic granularity is not a side effect but a core design goal.

4. Dual-level contrastive objectives

The representation-learning stage of DCCL is organized around instance-level and concept-level supervision. In the original framework, the total loss is

$$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$5

where $$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$6 is an instance-level contrastive term, $$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$7 is a conception-level contrastive term, and $$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$8 is a dispersion loss (Pu et al., 2023).

At the concept level, each sample is attracted to its current concept prototype and contrasted against the others: $$\mathcal{D}^U = \{ (x_i^U, y_i^U) \}_{i=1}^{M^U} \subset \mathcal{X} \times \mathcal{Y}^U,$$9 This loss turns the dynamically mined concept assignment $\mathcal{Y}^L \subset \mathcal{Y}^U$0 into a prototype-level positive and all other prototypes into negatives (Pu et al., 2023).

The dispersion loss imposes an inter-conception margin within a concept-conditioned mini-batch. If $\mathcal{Y}^L \subset \mathcal{Y}^U$1 denotes instances from concept $\mathcal{Y}^L \subset \mathcal{Y}^U$2 in a concept batch, then

$\mathcal{Y}^L \subset \mathcal{Y}^U$3

and

$\mathcal{Y}^L \subset \mathcal{Y}^U$4

The prototype memory is updated online by

$\mathcal{Y}^L \subset \mathcal{Y}^U$5

The paper describes this as an EMA-like mechanism for consistent conception learning between periodic clustering updates (Pu et al., 2023).

The MGCE version keeps the concept-prototype contrast but reformulates the instance-level objective more explicitly. It combines self-supervised NT-Xent, supervised contrast on the labeled subset, and a switchable classification loss: $\mathcal{Y}^L \subset \mathcal{Y}^U$6 with a prototype-based concept loss

$\mathcal{Y}^L \subset \mathcal{Y}^U$7

and a single-expert objective

$\mathcal{Y}^L \subset \mathcal{Y}^U$8

When $\mathcal{Y}^L \subset \mathcal{Y}^U$9 is unknown, the classification branch reduces to supervised classification on labeled samples; when $N^L = |\mathcal{Y}^L|$0 is known, an unsupervised objective in the style of SimGCD is added (Zheng et al., 30 Sep 2025). This later formulation makes explicit that DCCL can be implemented as a dual-level representation learner even when parametric category counts are unavailable.

5. Architectures, training procedure, and empirical behavior

The original DCCL implementation uses a ViT-B/16 backbone pre-trained with DINO, with the [CLS] token as the feature representation $N^L = |\mathcal{Y}^L|$1 (Pu et al., 2023). The projection head $N^L = |\mathcal{Y}^L|$2 is an MLP used only for instance-level contrastive learning, while concept-level losses act directly on $N^L = |\mathcal{Y}^L|$3-normalized features. The method trains for 200 epochs, uses batch size 128, concept mini-batches with $N^L = |\mathcal{Y}^L|$4 concept labels and $N^L = |\mathcal{Y}^L|$5 instances per concept, and fine-tunes only the last ViT block at learning rate 0.01 while training the head at 0.1 with SGD with momentum and cosine annealing. Reported hyperparameters include $N^L = |\mathcal{Y}^L|$6, $N^L = |\mathcal{Y}^L|$7, $N^L = |\mathcal{Y}^L|$8, $N^L = |\mathcal{Y}^L|$9, $N^U = |\mathcal{Y}^U|$0, $N^U = |\mathcal{Y}^U|$1, $N^U = |\mathcal{Y}^U|$2 for fine-grained datasets and $N^U = |\mathcal{Y}^U|$3 for generic datasets, $N^U = |\mathcal{Y}^U|$4, and prototype momentum $N^U = |\mathcal{Y}^U|$5 (Pu et al., 2023).

On the standard GCD protocol, the method reports ACC on six datasets. For CIFAR-10, DCCL reaches 96.3 / 96.5 / 96.9 on All / Old / New, versus 91.5 / 97.9 / 88.2 for GCD. For CIFAR-100, it reports 75.3 / 76.8 / 70.2 versus 73.0 / 76.2 / 66.5. For ImageNet-100, it reports 80.5 / 90.5 / 76.2 versus 74.1 / 89.8 / 66.3 (Pu et al., 2023). The gains are especially marked on fine-grained benchmarks: on CUB-200, DCCL reports 63.5 / 60.8 / 64.9 versus 51.3 / 56.6 / 48.7, which the paper highlights as +16.2% on New; on Stanford Cars, 43.1 / 55.7 / 36.2 versus 39.0 / 57.6 / 29.9; and on Oxford-IIIT Pet, 88.1 / 88.2 / 88.0 versus 80.2 / 85.1 / 77.6 (Pu et al., 2023).

The ablation study isolates the contribution of each component. On CUB-200, the baseline instance-level model reports 51.3 / 56.6 / 48.7, then + Conception-level contrastive gives 54.9 / 52.3 / 55.4, + Momentum update of prototypes gives 57.7 / 54.0 / 58.1, + Dispersion loss gives 59.5 / 53.3 / 60.8, + Conception consolidation gives 60.1 / 59.4 / 60.7, and the Full DCCL gives 63.5 / 60.8 / 64.9. On Pet, the corresponding progression is 80.2 / 85.1 / 77.6, 81.6 / 80.7 / 81.0, 83.5 / 81.1 / 80.3, 84.3 / 83.1 / 84.5, 85.8 / 86.8 / 84.6, and 88.1 / 88.2 / 88.0 (Pu et al., 2023). The paper also reports that replacing DCG with semi-supervised k-means, even with the ground-truth number of clusters, underperforms DCG on CUB-200: 55.9 vs 63.5 (All) (Pu et al., 2023).

The later MGCE system retains a ViT-B/16 DINO encoder, fine-tunes only the last transformer block, uses 200 epochs, batch size 128, and a concept head $N^U = |\mathcal{Y}^U|$6 implemented as a 2-layer linear network (768 → 2048 → 768) (Zheng et al., 30 Sep 2025). Its default settings include $N^U = |\mathcal{Y}^U|$7 on fine-grained datasets, $N^U = |\mathcal{Y}^U|$8, $N^U = |\mathcal{Y}^U|$9, $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$0, $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$1, concept mini-batch size 128 with $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$2 and $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$3, graph threshold $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$4 on fine-grained data or 0.5 on generic data, multi-expert scaling $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$5, and loss weight $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$6 (Zheng et al., 30 Sep 2025). MGCE reports that it can automatically estimate the number of categories in unlabeled data and that, even without prior knowledge of category numbers, it outperforms parametric approaches that require knowing the exact number of categories, with an average improvement of 3.6\% across nine fine-grained benchmarks (Zheng et al., 30 Sep 2025).

6. Later developments, limitations, and broader interpretations

The main limitations identified in the original DCCL line are concentrated around concept estimation, graph construction, and scalability. The 2023 paper notes sensitivity to incorrect concept granularity, the need to tune $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$7 and $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$8, the possibility that spurious high-similarity edges in unlabeled data can mislead Infomap, the computational cost of building large similarity graphs even with FAISS, and the risk that an overly aggressive dispersion margin $$D_L = \{ (x_i^l, y_i^l) \}_{i=1}^N,\qquad D_U = \{ x_i^u \}_{i=1}^M,$$9 may hurt closely related fine-grained classes (Pu et al., 2023). The MGCE extension reports analogous sensitivities to $Y_L \subset Y_U$0 selection and $Y_L \subset Y_U$1, especially in early training when features are less discriminative, as well as reduced gains on generic datasets and spurious small clusters during dynamic mining (Zheng et al., 30 Sep 2025).

At the same time, later work broadens the interpretive scope of DCCL. The MGCE framework extends the single-expert paradigm to three experts at different information granularities, using

$Y_L \subset Y_U$2

and couples them with concept alignment matrices and a symmetric KL-based collaborative loss (Zheng et al., 30 Sep 2025). A plausible implication is that DCCL can be understood not only as a method for concept-conditioned contrastive learning, but also as a modular component in broader multi-granularity discovery systems.

A different broadening occurs in temporal networks. The paper "Contrastive Representation Learning for Dynamic Link Prediction in Temporal Networks" does not explicitly use the term DCCL, but its synthesis states that teneNCE operationalizes dynamic conceptual contrastive learning principles through local and global InfoNCE objectives over time (Nouranizadeh et al., 2024). In that interpretation, local conceptual alignment keeps the same node close to its future structural embedding, while global conceptual alignment keeps graph-level summaries coherent across time. This perspective treats DCCL less as a specific GCD algorithm than as a general design pattern for aligning latent concepts dynamically under evolving data.

The most persistent misconception is therefore not technical but terminological. DCCL is neither a single universally fixed acronym nor a single universally fixed algorithm. In its most established encyclopedia sense, it refers to the GCD family initiated by dynamic conception generation and prototype-based concept contrast (Pu et al., 2023) and later elaborated by multi-granularity concept mining and expert collaboration (Zheng et al., 30 Sep 2025). Beyond that core, the label also functions as a broader conceptual shorthand for methods that jointly update representations and latent concept structure under contrastive supervision.