Continual Learning & Catastrophic Forgetting
Characterizing and Reducing Pre-Trained ViT Dependence in Transfer-Based Methods
Generated automatically from the limitations stated in 20 papers (CVPR, NeurIPS, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
A principled account of when feature-frozen methods can be deployed outside the ImageNet-ViT regime, and design guidance for methods that degrade gracefully rather than collapse when pretraining is weak or absent.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Backbone-degradation ablation suite: take 3–5 representative feature-frozen methods (one adapter-based continual learning method, one geometric/frozen-feature method, one anomaly detection method) and re-evaluate them across a controlled pretraining gradient—ImageNet-21k, ImageNet-1k, small-scale supervised, self-supervised MAE, random init—holding the method fixed and measuring accuracy, forgetting, and AUROC as a function of backbone quality. Deliverable: a quantitative "degradation curve" per method showing where each breaks.
- 2
Cross-architecture transfer test: rerun the same methods on ConvNeXt, Swin, and ResNet backbones with matched pretraining data, measuring whether failures are transformer-specific or feature-quality-specific. Measured: rank correlation between backbone linear-probe accuracy and downstream method performance.
- 3
Domain-distance study: evaluate on domains far from ImageNet (medical, remote sensing, audio spectrograms) with domain-matched pretraining, to separate "the method needs good features" from "the method needs ImageNet features specifically."
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
The finding could be entirely unsurprising—everyone may already know performance tracks backbone quality, and the degradation curves may show smooth, predictable declines with no interesting structure or collapse points worth reporting.
Sub-problems
Continual Learning Under Backbone and Pre-Training Domain Shift
Current top-performing continual learning methods, especially prompt- and adapter-based approaches, rely almost entirely on frozen Vision Transformer backbones pre-trained on ImageNet. In real-world deployments where the target stream lies far outside natural image distributions (e.g., medical imaging, remote sensing) or where compute constraints necessitate lightweight non-ViT architectures, these methods lack pre-separated feature spaces and their performance guarantees evaporate. Consequently, the field cannot distinguish whether contemporary continual learning algorithms prevent catastrophic forgetting through robust plastic-stable dynamics or merely exploit the static separability of upstream pre-trained representations.
Anomaly Detection and Representation Adaptation Without Pre-Trained Foundation Backbones
Modern visual anomaly and out-of-distribution detection methods rely almost entirely on frozen features extracted from foundation models pre-trained on ImageNet. When applied to domains whose visual primitives are absent from standard pre-training datasets—such as specialized industrial inspection, semiconductor wafer maps, or specialized biomedical imaging—these methods degrade significantly or fail entirely. Because standard methods do not fine-tune backbones on normal-only target data due to feature collapse, practitioners in non-natural image domains are blocked from using state-of-the-art anomaly detection pipelines.
Evidence
Each paper's own statement of the limitation, verbatim.
- Representation-Steered Incremental Adapter-Tuning for Class-Incremental Learning with Pre-Trained ModelsCVPR 2026
Evaluated only on vision classification with ViT backbones; broader modalities are left to future work
- Exemplar-Free Class Incremental Learning via Preserving Class-Discriminative StructureCVPR 2026
Requires a high-quality pre-trained backbone (ViT) to provide the initial structured feature space upon which the geometric constraints rely.
- DK-DDIL: Adaptive Knowledge Retention for Dynamic Domain-Incremental Learning in Medical ImagingCVPR 2026
Evaluated only on image classification with ViT backbones; extension to cross-modal continual learning is left as future work
- SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental LearningCVPR 2025
Effectiveness depends on a pre-trained ViT whose feature space already separates base and novel classes reasonably; no evidence on domains far from ImageNet pre-training
- SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental LearningICLR 2025
Evaluated only on ViT backbones for image classification; the paper itself notes extension to other foundation models/architectures is untested
- Revisiting Pool-based Prompt Learning for Few-shot Class-incremental LearningICCV 2025
Requires a pre-trained ViT with VPT-Deep; effectiveness on non-ViT backbones or domains far from ImageNet is not shown
- Mixture of Noise for Pre-Trained Model-Based Class-Incremental LearningNeurIPS 2025
Requires ViT backbone pre-trained on ImageNet-21K; performance is untested on weaker or non-ViT backbones
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual LearningICLR 2025
Evaluation is limited to image classification with ViT features pre-trained on ImageNet; no evidence for other modalities or non-frozen backbones
Show all 20 papers
- LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual LearningCVPR 2025
Assumes a strong pre-trained ViT backbone and class-incremental, non-overlapping label spaces; evaluated only on ImageNet-R and CIFAR-100
- Intermediate Layer Classifiers for OOD generalizationICLR 2025
No fine-tuning of backbone weights is considered, so improvements are capped by the frozen pre-trained representations and do not address cases where no useful features exist at any layer
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly DetectionCVPR 2025
Requires strong self-supervised pretrained ViT backbones; performance tracks ImageNet linear-probing accuracy and MAE-pretrained models fail, so applicability depends on foundation model quality of the domain
- Continuous Subspace Optimization for Continual LearningNeurIPS 2025
Only evaluated on vision benchmarks (CIFAR100, ImageNet-R, DomainNet) with pre-trained ViTs; the authors themselves flag extension to multimodal settings as an open problem
- Achieving More with Less: Additive Prompt Tuning for Rehearsal-Free Class-Incremental LearningICCV 2025
Evaluated only on image classification with ViT-B/16 pretrained on ImageNet-21k; transfer to other architectures or modalities is not demonstrated
- Versatile Incremental Learning: Towards Class and Domain-Agnostic Incremental LearningECCV 2024
Evaluated only on ViT with adapters and image classification benchmarks (iDigits, CORe50, DomainNet); no other architectures or modalities tested
- Vector Quantization Prompting for Continual LearningNeurIPS 2024
Depends on a large pre-trained ViT backbone, inheriting pre-training data distribution biases and its computational cost
- F-OAL: Forward-only Online Analytic Learning with Fast Training and Low Memory Footprint in Class Incremental LearningNeurIPS 2024
Relies entirely on a powerful pre-trained encoder (ViT-B pretrained on ImageNet-1K); accuracy would collapse if no strong pre-trained backbone exists for the domain
- Convolutional Prompting meets Language Models for Continual LearningCVPR 2024
Requires a supervised ImageNet-21k pre-trained ViT-B/16 backbone; performance with weaker pretraining or non-transformer backbones is not evaluated
- Continuous Memory Representation for Anomaly DetectionECCV 2024
Requires a pre-trained backbone and normal-only training data; performance depends on features of the encoder network rather than end-to-end learned representations
- SimpleNet: A Simple Network for Image Anomaly Detection and LocalizationCVPR 2023
Dependence on pre-trained ImageNet backbones may limit effectiveness if the target domain features are not representable by the backbone.
- Generating Instance-level Prompts for Rehearsal-free Continual LearningICCV 2023
Requires ImageNet-pretrained ViT backbone; domain adaptation relies on pre-trained representations, and behavior on non-ViT or non-ImageNet backbones is not examined
Nearest existing work
- RanPAC: Random Projections and Pre-trained Models for Continual LearningNeurIPS 2023
- Dynamic Siamese Expansion Framework for Improving Robustness in Online Continual LearningNeurIPS 2025
- Proper Reuse of Image Classification Features Improves Object DetectionCVPR 2022
- Transfer without ForgettingECCV 2022
- PANDA: Adapting Pretrained Features for Anomaly Detection and SegmentationCVPR 2021
- Continual Normalization: Rethinking Batch Normalization for Online Continual LearningICLR 2022
- Learning Expandable and Adaptable Representations for Continual LearningNeurIPS 2025
- Continual Learners are Incremental Model GeneralizersICML 2023
- Select and Distill: Selective Dual-Teacher Knowledge Transfer for Continual Learning on Vision-Language ModelsECCV 2024
- ViT-EnsembleAttack: Augmenting Ensemble Models for Stronger Adversarial Transferability in Vision TransformersICCV 2025
- Adversarially robust transfer learningICLR 2020
- Preventing Catastrophic Forgetting through Memory Networks in Continuous DetectionECCV 2024
- Generalized Variational Continual LearningICLR 2021
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial TransferabilityNeurIPS 2025
- CHEEM: Continual Learning by Reuse, New, Adapt and Skip - A Hierarchical Exploration-Exploitation ApproachCVPR 2026
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
Sub-Quadratic Covariance Sketching for Exemplar-Free Continual Learning
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.
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.