- 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
- 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.
-
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.
-
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.
-
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]