---
title: Continual Category Discovery
url: https://www.emergentmind.com/topics/continual-category-discovery
type: topic
---

# Continual Category Discovery

Continual category discovery is a family of open-world learning problems in which a model is first trained on an initial labelled base set and is then repeatedly updated as new unlabelled data arrive over time. The defining requirement is dual: the learner must discover previously unseen semantic categories and must retain recognition of categories learned earlier, despite catastrophic forgetting and the absence, in many formulations, of stored past data. In the survey taxonomy, it is a temporal extension of category discovery that sits above the base settings of novel category discovery (NCD) and generalized category discovery (GCD), with the extra complication that the data distribution evolves across stages or timesteps [2509.22542]. Early work framed this setting as Continuous Category Discovery (CCD), while later literature introduced related formulations such as Continual Novel Class Discovery (CNCD), Continual Generalized Category Discovery (C-GCD), Generalized Continual Category Discovery (GCCD), rehearsal-free CCD, online continuous GCD, and on-the-fly category discovery, each emphasizing different assumptions about label-space overlap, memory, and streaming constraints [2210.04174][2405.06389][2507.17382][2308.12112][2503.09106][2408.13492][2604.11484].

## 1. Taxonomy and problem variants

The common CCD template has two stages. In the initial stage, a labelled dataset is used to learn an initial known class set; in the continuous stage, unlabelled data arrive sequentially, and each batch may contain known classes, novel classes, or both. GM introduced CCD explicitly in this form and emphasized that it is more realistic than static novel category discovery because data arrive continuously, unlabelled batches may contain both known and unknown classes, and the model must keep adapting without forgetting [2210.04174].

Several later formulations refine this template rather than replace it. CNCD, as used by FEA, assumes a base session with labelled known classes followed by \(T\) incremental discovery sessions with unlabelled data, with the convention that once a novel class is discovered, it becomes a known class in later sessions. In that formulation, the class sets are disjoint across sessions, so the current session’s unlabelled data correspond to the current novel set \(\mathcal{C}^t\), while previous novel classes move into \(\mathcal{C}^{0:t-1}\) for retention [2405.06389]. By contrast, C-GCD assumes that each incremental session contains a mixture of previously seen old classes and genuinely new classes, which makes separation, pseudo-labeling, and forgetting substantially harder [2507.17382].

The survey formalizes CCD as a derived setting of category discovery with several sub-scenarios: class-incremental CCD, mixed-incremental CCD, and self-supervised mixed-incremental CCD. It explicitly notes that the common class-incremental assumption, where each new batch contains only novel categories, is often unrealistic, and that mixed-incremental CCD is more representative because old and new classes coexist in the incoming unlabelled data [2509.22542]. This observation is echoed by C-GCD work such as Happy and GOAL, both of which adopt a rehearsal-free multi-session setting in which Stage-0 is labelled, later stages are unlabelled, and each later stage contains both old and new classes [2410.06535][2602.19872].

Two additional branches extend the field toward more restrictive deployment regimes. Rehearsal-Free Continual Category Discovery (RF-CCD) forbids replaying stored data and, in the FAC formulation, assumes disjoint class sets across sessions while using strong pretrained backbones and classifier-only updates in the unlabelled stages [2503.09106]. Online and on-the-fly formulations such as OCGCD, DEAN, and PACO move from session-level updates toward streaming inference, where the model must process batches or even individual samples as they arrive, cannot reorganize the whole stream offline, and must make sequential known/novel/attach/create decisions under evolving prototype memory [2408.13492][2604.11484].

## 2. Formal setup and evaluation protocols

A canonical mathematical setup starts with a labelled base dataset
\[
\mathcal{D}_{train}^0=\{(\mathbf{x}_i^0,y_i^0)\}_{i=1}^{N^0},
\]
drawn from base categories, followed by a sequence of later datasets
\[
\{\mathcal{D}_{train}^t\}_{t=1}^T,
\]
on which the model is updated at each timestep \(t\) to identify novel categories while preserving knowledge from earlier steps [2509.22542]. GM instantiates this as an initial labelled dataset \(\mathcal{D}_{train}^0\) over \(\mathcal{C}^0\), followed by sequential unlabelled datasets \(\mathcal{D}_{train}^t\) that may contain both known classes seen before and novel classes not yet discovered [2210.04174]. GOAL adopts the same broad setting but focuses on C-GCD: a small labelled base set is followed by later sessions containing unlabelled data drawn from a mixture of old classes and previously unseen classes [2602.19872].

Evaluation in continual category discovery is more heterogeneous than in static GCD because methods must measure both discovery and forgetting. GM introduced maximum forgetting \(\mathcal{M}_f\), defined as the maximum drop in known-class accuracy over time relative to the initial accuracy, and final discovery \(\mathcal{M}_d=\mathrm{ACC}_{\mathrm{novel}^T}\), together with \(\mathrm{ACC}_{\mathrm{known}^t}\) and \(\mathrm{ACC}_{\mathrm{novel}^t}\) under Hungarian matching [2210.04174]. Happy redefined \(\mathcal{M}_f\) as
\[
\mathcal{M}_f = \max_t \{ACC_\text{init}^0-ACC_\text{init}^t\},
\]
and \(\mathcal{M}_d\) as
\[
\mathcal{M}_d = \frac{1}{T}\sum_{t=1}^T ACC_\text{new}^t,
\]
while reporting “All,” “Old,” and “New” accuracy with Hungarian matching over all classes seen so far [2410.06535]. GOAL follows the same “All,” “Old,” and “New” reporting and specifies inductive evaluation, with predictions matched to ground truth via the Hungarian algorithm over all classes seen so far [2602.19872].

PromptCCD argues that single-step GCD accuracy is insufficient in a continual setting and extends the usual metric to continual ACC (cACC). Its cACC sequentially evaluates each stage, appends the pseudo-labelled samples to the labelled pool for future stages, and averages ACC over stages, thus making early-stage errors persistent in the metric itself [2407.19001]. The survey adopts cACC as the main CCD metric, again reported on All, Old, and New, and notes that this cumulative perspective is more faithful to continuous deployment than single-step evaluation [2509.22542].

Open-world variants add still other metrics. OpenGCD introduces harmonic clustering accuracy (HCA), which combines accuracy on known classes with clustering ACC on novel classes through a harmonic form, to avoid the misleading symmetry of ordinary ACC in a generalized discovery pipeline [2308.06926]. This suggests that evaluation remains an active design choice rather than a settled convention.

## 3. Fundamental technical tensions

A recurrent claim across the literature is that continual category discovery is difficult not merely because of catastrophic forgetting, but because discovery and recognition impose incompatible representational pressures. GM states the conflict directly: classification on known classes prefers stable, discriminative, compact features, whereas discovery of new classes prefers rich, diverse, and more flexible features so that unseen categories can separate and cluster well [2210.04174]. FEA recasts the same problem as a feature-discrepancy issue: supervised learning on known classes yields class-discriminative features suited for classification, but novel-class discovery needs more diverse and rich features; it also identifies inter-session confusion, where classes trained in different sessions produce unstable decision boundaries in a shared feature space [2405.06389].

C-GCD work sharpened this diagnosis. Happy identifies prediction bias and hardness bias as two distinct failure modes. Prediction bias arises because old classes are trained with full supervision at Stage-0 while new classes are learned later through unsupervised self-training, creating a confidence gap that pulls new samples toward old classes. Hardness bias arises because some old classes are intrinsically more confusable than others; if replay or prototype sampling is uniform, the difficult classes are underrepresented and forgotten unevenly [2410.06535]. GOAL attributes poor long-horizon behavior in prior C-GCD systems to prototype drift and inconsistent feature alignment: if the classifier is repeatedly updated, earlier knowledge can be overwritten, and if the feature space is not constrained by a fixed global geometry, the alignment target shifts across sessions and produces category confusion, especially for novel classes that lie close to known ones [2602.19872].

A more distributional interpretation appears in VB-CGCD. That work argues that forgetting is fundamentally tied to class-distribution mismatch, especially covariance misalignment between old and new classes. Each class is modeled as a Gaussian, and the claim is that as variational inference fits a new class, its covariance may grow in a way that shifts decision boundaries irreversibly and degrades old-class performance [2507.17382]. This is a different explanation from pure feature-drift accounts, but it points to the same operational problem: continual discovery destabilizes the geometry inherited from previous sessions.

Other formulations expose additional failure modes. NCENet focuses on representation learning and argues that old knowledge should be preserved at the level of relational structure rather than only logits, motivating contrastive knowledge distillation [2412.05573]. Virtual Category-Guided C-GCD emphasizes ambiguous unlabelled samples that cannot be confidently assigned to known classes nor reliably grouped as novel ones, making pseudo-labeling brittle and biasing learning toward familiar categories [2607.04984]. PACO extends this to streaming inference, arguing that a single static threshold is fundamentally flawed because the model must continuously decide whether a sample belongs to a known class, matches an existing novel category, or should initiate a new one; the confidence landscape changes as the novel memory bank matures [2604.11484].

## 4. Methodological families

The field has produced several design families that can be read as different answers to the stability–plasticity conflict.

One influential family separates discovery and consolidation into different phases or branches. GM alternates a growing phase and a merging phase. The growing phase expands feature diversity through continuous self-supervised learning, detects likely novel samples by thresholding distance to known prototypes, and learns a cluster head from WTA-hashing pairwise similarity. The merging phase then sifts noisy pseudo-labels using local density, applies pseudo label learning on exemplars and prototypes, distills a frozen static branch into a dynamic branch, and fuses branches with exponential moving average [2210.04174]. CAMP also separates representation adaptation from classifier stabilization: it combines self-supervised and supervised contrastive learning with a learnable projector for feature distillation, then corrects prototype drift through a Prototype Adaptation Network that maps old centroids into the new latent space [2308.12112].

A second family treats prototypes, proxies, or prompt banks as the core memory structure. Early proxy-anchor CGCD methods fine-tuned a feature extractor and proxy anchors on the labelled set, split unlabelled data into old and novel subsets, clustered novel samples with affinity propagation, and mitigated forgetting through proxy-anchor-based exemplar replay plus feature distillation [2307.10943]. CATEGORIZER extends this line with Extreme Value Theory (EVT): it models inclusion boundaries around proxy anchors with Weibull distributions, rejects low-probability samples as unknown, clusters the unknown set with affinity propagation, prunes redundant proxies through EVT-based model reduction, and adds replay and distillation in the continual stage [2504.08550]. IDOD replaces distillation-heavy memory with a static pool of representative features, a dynamic pool of accumulated novel representations, and mutually orthogonal prototypes that anchor classification while keeping storage overhead low [2508.11173]. PromptCCD replaces a fixed prompt pool with a Gaussian Mixture Prompting module, so that prompts are distributional prototypes whose means act both as prompts and as a parametric memory of discovered categories; the split-and-merge GMM also gives online class-number estimation [2407.19001].

A third family emphasizes neighborhood structure, self-distillation, and meta-learning. MetaGCD simulates future incremental episodes during offline training and optimizes a bi-level meta-objective so that the initialization is explicitly good at later discovery and retention. During adaptation it uses a soft neighborhood-based contrastive loss in which nearby samples contribute as soft positives with learned weights rather than binary assignments [2308.11063]. NCENet constructs \(k\)-nearest-neighbor neighborhoods inside each mini-batch, averages them into local commonalities, and uses those commonalities as a semantic basis for self-distillation in its Neighborhood Commonality-aware Representation Learning module; old knowledge is preserved through Bi-level Contrastive Knowledge Distillation, with both student-anchored and teacher-anchored contrastive losses [2412.05573]. Virtual Category-Guided C-GCD similarly attacks pseudo-label brittleness by routing uncertain samples into temporary virtual categories and combining that mechanism with Expanded Neighborhood Contrastive Learning over neighbors of neighbors [2607.04984].

A fourth family focuses on single-head or distribution-guided objectives rather than auxiliary cluster heads. FEA explicitly criticizes the multi-head paradigm and proposes a unified cosine classifier trained with a guide-to-novel term, a centroid-to-samples similarity constraint (CSS), and a boundary-aware prototype constraint (BAP). Old knowledge is preserved with prototype-based pseudo rehearsal and feature distillation, while novel discovery is driven by prior-distribution alignment rather than one-hot pseudo-labels [2405.06389]. Happy also uses a single unified classifier but attacks bias directly through clustering-guided initialization, soft entropy regularization that balances probability mass between old and new groups, hardness-aware prototype sampling, and feature-space Gaussian replay [2410.06535].

A fifth family seeks a stable global geometry. GOAL is the clearest example. It replaces drifting session-wise classifier prototypes with a fixed Equiangular Tight Frame (ETF) classifier inspired by Neural Collapse. For \(K\) classes, the ETF prototypes satisfy
\[
p_k^\top p_j=
\begin{cases}
1, & k=j,\\
-\frac{1}{K-1}, & k\neq j,
\end{cases}
\]
so all classes are equally separated angularly. GOAL freezes this ETF classifier across the whole continual process, performs supervised alignment for labelled samples in the base session and confidence-guided alignment for novel samples in incremental sessions, and thereby preserves old knowledge through a fixed geometric target rather than memory replay [2602.19872].

A sixth family adopts explicitly probabilistic or frozen-representation views. VB-CGCD models each class as a variational Gaussian, uses covariance-aware nearest-class-mean classification with Mahalanobis geometry, applies self-correcting relabeling, and stops optimization early when the determinant scale of new-class covariance aligns with old covariances [2507.17382]. FAC goes in the opposite direction and argues that, with strong pretrained backbones, continual updates on unlabelled data do not reliably improve representations and may even degrade them; its baseline therefore freezes the backbone after the supervised base stage, clusters each new session with \(k\)-means, estimates unknown class counts by over-clustering and greedy merging, and trains only the classifier using Gaussian-replayed cluster statistics and logit normalization [2503.09106].

## 5. Empirical benchmarks and representative results

The empirical literature is benchmark-driven. Four datasets recur across many C-GCD papers: CIFAR-100, TinyImageNet, ImageNet-100, and CUB or CUB-200 [2602.19872][2507.17382][2410.06535]. Broader CCD evaluations also include Caltech-101, FGVC-Aircraft, Stanford Cars, and the staged iNatIGCD benchmark highlighted in the survey as especially valuable because it includes temporal progression, geographic or domain shift, fine-grained species recognition, and realistic incremental structure [2407.19001][2509.22542]. Fine-grained datasets such as CUB-200, MIT67, Stanford Dogs, and FGVC-Aircraft are also common in proxy-based and diversity–discrimination work [2307.10943][2508.11173].

Quantitatively, the recent trend is toward stronger long-horizon and lower-label protocols. GOAL reports that, compared with Happy, it reduces average forgetting by 16.10% and improves novel-category discovery by 3.19% across CIFAR100, TinyImageNet, ImageNet-100, and CUB; in a 10-stage setting, the gains become larger, with a 19.14% forgetting reduction and a 10.28% novel-class accuracy improvement on CIFAR100, and a 23.71% forgetting reduction and a 4.19% new-class accuracy improvement on TinyImageNet [2602.19872]. VB-CGCD reports a +15.21% average gain in final-session overall accuracy over prior methods on standard benchmarks and introduces a b10t9 setting with only 10% labelled data and 9 continual sessions, where it reaches 67.86% final accuracy compared with 38.55% for the prior state of the art reported in the paper [2507.17382].

Prompt- and representation-centric methods have also posted strong results. PromptCCD reports cACC gains over Grow and Merge, MetaGCD, PA-CGCD, ORCA, GCD, and SimGCD on several datasets; under known class counts with DINO, it reports 64.17% cACC on CIFAR100, 76.16% on ImageNet-100, 61.84% on TinyImageNet, and 55.45% on CUB, with further gains under DINOv2 [2407.19001]. FEA reports that it consistently outperforms baselines, especially as the number of sessions increases, including gains over FRoST on Tiny-ImageNet of 4.17%, 5.11%, and 7.92% average accuracy for \(T=2,5,10\), respectively [2405.06389]. NCENet reports that, in the last incremental learning session on CIFAR100, its clustering accuracy exceeds the second-best method by 3.09% on old classes and by 6.32% on new classes [2412.05573].

Online and rehearsal-free variants show a similar push toward realistic deployment. PACO, which is formulated for on-the-fly discovery with one-pass streaming inference, reports mean performance over five runs of 57.76 All, 72.14 Old, and 48.07 New under Strict-Hungarian, and 65.16 All, 84.41 Old, and 52.84 New under Greedy-Hungarian, while also improving AGE and PHE-12bit as a plug-and-play inference module [2604.11484]. FAC reports strong gains across rehearsal-free benchmarks, with final Last Acc values of 66.2 on CUB200, 35.6 on Stanford Cars196, 39.5 on iNat550, and 73.7 on Tiny-ImageNet200, outperforming both RF-CCD-native and combined CL+NCD baselines in its study [2503.09106]. DEAN, in the OCGCD setting, reports major gains over FRoST and CGCD on CUB200; for the configuration \([160(0.8) / 160(0.2) + 40(1.0)]\), it reports \(M_{all}=59.56\), \(M_{old}=66.33\), \(M_{new}=33.11\), and \(F=16.26\), compared with much lower values for the baselines [2408.13492].

A plausible implication of these results is that the field has shifted from demonstrating feasibility on short incremental sequences toward comparing memory structures, geometry, and class-estimation mechanisms under longer horizons, smaller labelled bases, and more realistic mixed old–new streams.

## 6. Open problems, assumptions, and research directions

Despite rapid progress, the literature repeatedly points to unresolved assumptions. Class-number estimation remains one of the hardest components. GM can estimate the number of novel clusters using semi-\(k\)-means on mixed exemplars and incoming data, with only a slight performance drop [2210.04174]. PromptCCD uses a split-and-merge GMM; Happy uses the silhouette score; PA-CGCD and DEAN rely on non-parametric clustering such as affinity propagation; FAC estimates the class count by over-clustering and greedy merging; GOAL explicitly notes that it still assumes the number of novel classes is given or estimated; and the survey treats continual class-number estimation as intertwined with retention, since overestimating causes fragmentation and underestimating causes class merging and forgetting [2407.19001][2410.06535][2307.10943][2408.13492][2503.09106][2602.19872][2509.22542].

Memory and storage assumptions are equally unsettled. Several methods are rehearsal-free by design, often for privacy or storage reasons, but replace raw-data replay with prototypes, pseudo-features, Gaussian statistics, prompts, or virtual categories [2405.06389][2410.06535][2407.19001][2607.04984]. VB-CGCD notes the quadratic storage cost of full covariance matrices and the continuing difficulty of small-sample covariance estimation, especially on datasets such as CUB [2507.17382]. IDOD argues that representative representation replay can reduce extra storage dramatically relative to replay-heavy baselines, reporting only 3.5M–5.4M extra storage versus much larger costs for GM, PA-CGCD, and PromptCCD* in its comparison [2508.11173].

Another recurring limitation is the heavy reliance on large self-supervised backbones. The survey repeatedly notes that large pretrained backbones such as DINO improve results substantially and should be regarded as the default rather than training from scratch [2509.22542]. FAC pushes this claim further by arguing that, in rehearsal-free settings with strong pretrained models, representation learning during the unlabelled stages may not improve features and may even degrade them, which motivates freezing the backbone rather than continually fine-tuning it [2503.09106]. This claim does not invalidate continual adaptation, but it sharpens a methodological controversy: whether continual discovery should primarily update representations, update classifiers and memories, or stabilize a fixed geometric or probabilistic structure over time.

The literature also points toward broader deployment scenarios. GOAL suggests adaptive expansion of the fixed ETF prototype set and multimodal discovery with text or semantic cues [2602.19872]. The survey calls for unified CCD frameworks that jointly address continual learning, class discovery, imbalance, and domain shift; adaptive class-number inference during training; better memory and consolidation mechanisms; more realistic benchmarks; hybrid label-assignment strategies; cross-modal and semantic cues; and a move beyond simple image-level scenes toward instance-level and multi-object environments [2509.22542]. OpenGCD makes a related operational point: in a full open-world system, category discovery is not an isolated clustering subroutine but one stage in a pipeline linking open-set recognition, grouped human verification, and incremental learning [2308.06926].

Continual category discovery has therefore evolved into a broad research program rather than a single benchmark task. Its central question remains stable across formulations: how to absorb genuinely new semantic structure from unlabelled streams without destroying the decision structure already learned. What has changed is the level at which answers are proposed—geometry, probability, prompts, proxies, neighborhoods, virtual categories, or frozen representations—and the field now increasingly evaluates those answers under mixed streams, long horizons, limited labels, and online decision constraints [2602.19872][2507.17382][2604.11484].

Source: https://www.emergentmind.com/topics/continual-category-discovery