acceptodds
Under review as a conference paper at ICLR 2027

CDC-OIL: Class-Domain Coupled Online Incremental Learning

Abstract

Real-world data streams often exhibit simultaneous class increment and domain shift, while existing continual learning settings typically study these two sources of non-stationarity in isolation. So we introduce Class-Domain Coupled Online Incremental Learning (CDC-OIL), where class and domain spaces evolve jointly or separately under single-pass data access and ambiguous task boundaries. We further identify a bidirectional interaction between class increment and domain shift. Class increment encourages model to refine semantic features but compress the feature space, increasing sensitivity to domain shift. Conversely, domain shift can enrich cross-domain semantics, but excessive shifts distort class distributions and hinder new-class discrimination. To address this challenge, we propose Class-Domain Decoupled Extensible Gaussian Mixture Models (CDD-EGMMs), which factorize feature distributions into class-specific semantic mixtures and domain shifts shared across classes. This compositional memory supports online updates and generative replay by recombining semantic and domain components without storing raw data. A Gaussian Representation Constraint further promotes class separability as the label space expands. Experiments on four benchmarks demonstrate that CDD-EGMMs consistently improves accuracy and reduces forgetting compared with representative continual learning baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Data stream, continual learning, domain shift, class incremental learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PjTZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Class-Domain Coupled Online Incremental Learning, a continual-learning setting in which class expansion and domain shift occur jointly or independently under single-pass data access and unknown task boundaries. Its main contribution is Class-Domain Decoupled Extensible Gaussian Mixture Models, which factorize feature distributions into class-specific semantic Gaussian mixtures and domain-specific shifts represented by a shared low-rank basis. The authors report improved average accuracy and reduced forgetting over several continual-learning baselines.

Critical strengths

The paper addresses a realistic limitation of standard continual-learning benchmarks: class expansion and domain shift often occur simultaneously rather than in isolation.

Critical weaknesses
  1. The proposed setting appears to combine several existing constraints: class increment, domain increment, online single-pass learning, and unknown task boundaries. This may still be a useful benchmark. However, some discussion should be added:
  • which assumptions are genuinely new;
  • which prior methods already satisfy them;
  • whether the proposed setting changes the optimization problem or mainly changes the data protocol.
  1. The Gaussian memory is updated using detached features, while the feature extractor continues to change through training on real incoming data. Replay features are used only to train the classifier, not the feature extractor. Consequently, the stored Gaussian distributions may become inconsistent with the current representation space.

  2. The method maintains K Gaussian components per class, each with a full covariance matrix, together with domain states and a shared basis. Its memory complexity can grow substantially with the number of classes, feature dimensions, and domains. The experiments are conducted on CIFAR-10/CIFAR-100, DomainNet, and CORe50. While DomainNet is more challenging, the overall evaluation remains relatively small compared with modern large-scale vision benchmarks, and does not include ImageNet-1000, stronger backbone ResNet-50, ViT.

  3. The method requires domain labels after each prediction to select update branches and maintain domain states. This is acceptable as a supervised benchmark assumption, but it is stronger than what is available in many real streaming applications.

Questions for authors

Refer to [Weakness]

Reviewer_hELoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Class-Domain Coupled Online Incremental Learning (CDC-OIL), a setting where both classes and domains can evolve within a single-pass stream. It also proposes CDD-EGMMs which models visual features using class-specific gaussian mixtures and domain-shifts. It generates synthetic feature replay for training classifier. Experiments use CIFAR , DomaiNet and Core50 datasets.

Critical strengths
  • Class expansions and domain shifts ocurring together with online setting is very realistic and a relevant problem setting.
  • Paper compares with many online CL baselines and also reports five-seed results and ablations
Critical weaknesses
  • The claims about generality and distinction from existing settings needs more justification. CDC-OIL is presented as a realistic online setting without task identities or boundaries, but there is dependency on domain labels. This is an important assumption as domain changes maybe gradual or overlapping or unknown? and many Domain-IL methods also do not explicitly use domain labels.
  • The class-domain factorization relies on a strong assumption. The separation of visual representation to mu (domain-invariant semantic feature) and A (cross-class domain shift term) needs more clarification. Because Domain shifts can interact differently with different classes, Ex. illumination changes might affect visually similar class differently, or style change can modify class-specific features. Also, can these components be reliably estimated from single-pass data.
  • The paper comparisons are difficult to assess because none of the method were designed for this new CDC-OIL protocol. It might need comparison with general incremental learning
  • The theoretical analysis also relies on idealized assumptions
Questions for authors
  • What evidence supports a shared additive domain shift across classes? How can the class semantics and domain shift be reliably separated, especially when a new class and a new domain appear simultaneously?
  • How realistic are the generated replay features?
  • CDD-EGMMS stores class-conditional gaussian means and covariances in feature space of encoder. But encoder changes as new samples arise, and wont this create a mismatch? How are the stored Gaussian distributions aligned with the changing feature space.
  • Need more details on the resources and memory efficiency
  • Can the authors provide more directly matched comparisons
  • There are multiple components making contribution and interaction with others difficult to isolate
Reviewer_KjRuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces Class-Domain Coupled Online Incremental Learning (CDC-OIL), where classes and domains can evolve independently or simultaneously. To address this setting, the authors propose a method that decouples class-specific Gaussian mixtures from shared domain shifts, with feature-level replay and a representation constraint to mitigate forgetting. The authors report improved accuracy and reduced forgetting over existing continual learning baselines across four benchmarks.

Critical strengths

The paper proposes a method specifically designed to handle the considered class-domain incremental learning setting.

Critical weaknesses
  1. Unclear problem novelty and unconvincing experimental validation. The proposed CDC-OIL setting appears to overlap with Versatile Incremental Learning (Park et al., ECCV 2024), and a clearer distinction would help establish its novelty. The related work discussion could also be improved, particularly regarding XIL (Choi et al., ICLR 2026), which focuses on bidirectional domain transferability. Despite emphasizing online learning, the paper provides limited discussion of relevant online continual learning settings and does not report online accuracy. Moreover, most baselines do not directly address joint class-domain evolution, while relevant methods such as ICON (Park et al., ECCV 2024) are omitted. Addressing these points would help strengthen the experimental evaluation.
  2. Limited methodological novelty and insufficient theoretical justification. The proposed method combines existing techniques, including GMMs, additive domain decomposition, feature replay, and margin-based regularization, but its distinctive technical contributions could be more clearly highlighted. Furthermore, the theoretical analysis assumes an additive class-domain decomposition, whose validity and learnability would benefit from further justification. Additional empirical evidence would also help substantiate the claimed benefits of class-domain decoupling.
Questions for authors

I would appreciate further clarification of the paper's positioning relative to existing work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.