Continual Learning & Catastrophic Forgetting
Sub-Quadratic Covariance Sketching for Exemplar-Free Continual Learning
Generated automatically from the limitations stated in 6 papers (CVPR, ICML, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Exemplar-free continual learning methods increasingly rely on storing per-class feature statistics (prototypes and full covariance matrices) to generate pseudo-replay features or parameterize Mahalanobis classifiers without retaining raw inputs. However, storing a dense $D \times D$ covariance matrix for each of $C$ classes incurs $\mathcal{O}(C \cdot D^2)$ storage, which quickly exceeds device memory when scaling to large label spaces ($C \ge 1000$) or modern high-dimensional embeddings ($D \ge 1024$). Practitioners currently resort to naive diagonal approximations that discard inter-feature correlations and cause severe classification degradation under continual distribution shifts. As a result, covariance-based exemplar-free learning is structurally blocked from scaling to realistic large-vocabulary benchmarks on memory-constrained hardware.
Why it matters
Enables exemplar-free and statistical-replay continual learning to scale to tens of thousands of classes and high-dimensional foundation model backbones on memory-constrained edge hardware.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Low-Rank Plus Shared-Basis Covariance Decomposition: Model class covariances as a rank-$k$ class-specific factor ($k \ll D$) over a shared global subspace plus diagonal residual noise; measure total parameter storage (in megabytes) against incremental accuracy on Split ImageNet-1k.
- 2
Streaming Frequent Directions for Class Statistics: Adapt deterministic streaming sketching to update compact low-rank covariance sketches online as each class stream arrives; measure relative spectral norm error ($\|\Sigma_c - \hat{\Sigma}_c\|_2$) and task-retention performance relative to full-covariance baselines.
- 3
Kronecker / Block-Structured Covariance Approximations: Exploit layer-wise or channel-group structures in pretrained Vision Transformer embeddings to store class covariances as compact Kronecker products; measure classification accuracy and peak memory across ViT and ConvNeXt feature backbones.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If frozen global covariance matrices or fixed feature extractors from large foundation models reduce intra-class covariance variations to negligible noise, class-specific covariance storage would become entirely redundant.
Evidence
Each paper's own statement of the limitation, verbatim.
- Representation-Steered Incremental Adapter-Tuning for Class-Incremental Learning with Pre-Trained ModelsCVPR 2026
Requires storing frozen per-class prototype and covariance statistics, whose memory grows with the number of classes
- Learning to Memorize with Attributive and Associative Memory for Online Test-Time Adaptation of Vision-Language ModelsICML 2026
Maintains C separate D×D memory matrices, so memory/compute scales with class count, potentially prohibitive for very large label spaces
- Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction for Exemplar-Free Class-Incremental LearningICML 2025
Requires storing per-class covariance matrices (d^2 per class), which scales quadratically with feature dimension and may be large for high-dimensional backbones
- Semantically-Shifted Incremental Adapter-Tuning is A Continual ViTransformerCVPR 2024
Requires storing per-class prototypes and covariance matrices whose size grows with the number of classes learned
- Long-Tail Class Incremental Learning via Independent Sub-prototype ConstructionCVPR 2024
Storing per-class mean and full covariance matrices plus correlation matrices scales with the number of classes, so the 'no replay data' memory claim still incurs growing storage
- Class-Incremental Learning via Dual AugmentationNeurIPS 2021
Requires storing class-wise covariance matrices, which leads to memory overhead that scales quadratically with feature dimensions unless diagonal approximations are used.
Nearest existing work
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental LearningNeurIPS 2024
- REMIND Your Neural Network to Prevent Catastrophic ForgettingECCV 2020
- Tensor Decomposition Based Memory-Efficient Incremental LearningICML 2025
- Forget Forgetting: Continual Learning in a World of Abundant MemoryICLR 2026
- Exemplar-free Continual Representation Learning via Learnable Drift CompensationECCV 2024
- Constrained Few-Shot Class-Incremental LearningCVPR 2022
- Computationally Budgeted Continual Learning: What Does Matter?CVPR 2023
- EEC: Learning to Encode and Regenerate Images for Continual LearningICLR 2021
- Online Task-Free Continual Learning via Dynamic Expansionable Memory DistributionCVPR 2025
- Resurrecting Old Classes with New Data for Exemplar-Free Continual LearningCVPR 2024
- Gradient-based Editing of Memory Examples for Online Task-free Continual LearningNeurIPS 2021
- FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual LearningNeurIPS 2023
- Online Continual Learning with Maximal Interfered RetrievalNeurIPS 2019
- CRAM: Large Scale Video Continual Learning with Bootstrapped CompressionICCV 2025
- Continual Learning in Low-rank Orthogonal SubspacesNeurIPS 2020
Related open problems
Continual Learning & Catastrophic Forgetting
A Standardized Transfer-Validation Benchmark for Methods Validated Only on Small-Scale Image Classification
Nearly every training-time method in the modern literature—continual learning, distillation, unlearning, robustness, federated aggregation—is validated exclusively on CIFAR/ImageNet-subset image classification with ResNet-18/50-class backbones, and each paper explicitly defers scale, architecture, and modality transfer to future work. Because every paper defers this simultaneously, the field has no evidence base at all for whether any of these methods survive the shift to transformers, higher resolutions, dense tasks, or non-vision data; practitioners choosing between methods are choosing based on measurements made in a regime none of them deploy in. The barrier is not one method's limitation but a collective, unpriced evaluation debt: nobody can tell which small-scale conclusions are regime-robust and which are artifacts of ResNet-18 on 32×32 images.
Continual Learning & Catastrophic Forgetting
Characterizing and Reducing Pre-Trained ViT Dependence in Transfer-Based Methods
A wide range of methods built on frozen ImageNet-pretrained ViT features—whether for adapting to new tasks, constraining representations, or detecting anomalies—inherit an unexamined dependency: their guarantees hold only when the backbone's feature space is already good. Today nobody knows whether these methods degrade gracefully with weaker pretraining, generalize to domains far from ImageNet, or transfer to non-transformer architectures, because the dependence is never ablated. The consequence is that reported gains may be properties of the backbone, not of the methods themselves, and the methods are silently inapplicable wherever a strong ImageNet ViT does not exist.
Continual Learning & Catastrophic Forgetting
Task-Agnostic Inference for Modular and Masked Continual Learning Architectures
Modular continual learning methods prevent catastrophic forgetting by allocating task-specific masks, sub-networks, or adaptive parameters during training. However, these methods structurally depend on having ground-truth task identifiers provided at test time to select the corresponding parameters. In practical deployments, incoming inputs arrive without task metadata, rendering task-dependent parameter masking unusable in class-incremental or task-agnostic settings. Without a reliable mechanism to resolve task identity or route inputs dynamically at test time, parameter-isolation architectures remain restricted to artificial task-incremental benchmarks.