Research map / Continual learning / Class-incremental
Dynamic Integration of Task-Specific Adapters for Class Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DIA, a non-exemplar class-incremental learning framework that integrates patch-level task-specific adapters via signature vectors and aligns models with patch-level distillation and patch-level feature reconstruction.
Problem
Non-exemplar class incremental learning (NECIL) with pre-trained ViTs suffers catastrophic forgetting and task interference, with feature drift of old samples and inaccurate classifier decision boundaries when no old exemplars can be stored.
Approach
Task-specific bottleneck adapters per block are routed per patch token through learned signature vectors and softmax-merged, while distillation weights patch tokens by their contribution to new-task learning and old-class features are reconstructed from patch tokens plus prototypes instead of Gaussian sampling.
Key results
- DIA reaches 85.61% average accuracy on ImageNet-R (10 tasks), 71.58% on ImageNet-A, 93.21% on CUB-200 and 94.29% on CIFAR-100, reported as SOTA.
- DIA surpasses EASE by 3.88%, 6.24%, 1.62%, and 1.94% on ImageNet-R, ImageNet-A, CUB-200, and CIFAR-100 respectively, with about 90% fewer FLOPs than EASE.
- With 20 tasks, DIA achieves 83.51% average accuracy on ImageNet-R, outperforming the second-best by 2.05%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the complete DIA framework rather than the learnable-classifier pretrained encoder (ablation) | ViT-B/16 · ImageNet-R, 10 tasks | Average accuracy: 40.35 -> 85.61 | Helped |
| Add PFR to TSAI and PDL (ablation) | DIA · ImageNet-R, 10 tasks | Average accuracy: 84.55 -> 85.61 | Helped |
| Add PDL to TSAI and PFR (ablation) | DIA · Cifar-100, 10 tasks | A10 accuracy: 89.81 -> 90.80 | Helped |
| Use angular similarity in PDL instead of Euclidean distance (ablation) | DIA · ImageNet-R, 10 tasks | Average accuracy: 81.73 -> 85.61 | Helped |
| Use angular similarity in PDL instead of cosine similarity (ablation) | DIA · Cifar-100, 10 tasks | A10 accuracy: 89.67 -> 90.80 | Helped |
| Apply PDL to patch tokens only, not the class token (ablation) | DIA · ImageNet-R, 10 tasks | Average accuracy: 83.87 -> 85.61 | Helped |
| Use patch-level PDL instead of feature-drift loss (ablation) | DIA · Cifar-100, 10 tasks | Average accuracy: 92.97 -> 94.29 | Helped |
| Use PFR instead of Gaussian feature reconstruction (ablation) | DIA · ImageNet-R, 10 tasks | Average accuracy: 83.72 -> 85.61 | Helped |
| Use PFR instead of SLCA feature reconstruction (ablation) | DIA · Cifar-100, 10 tasks | Average accuracy: 93.79 -> 94.29 | Helped |
| Increase the PDL weight lambda (ablation) | DIA · Cifar-100 | Accuracy | Hurt |
| Use small beta for prototype-feature reconstruction (ablation) | DIA · Cifar-100 | Accuracy | Hurt |
| Add DIA instead of CPrompt | ViT-B/16 · ImageNet-R, 10 tasks | Average accuracy: 82.92 -> 85.61 | Helped |
Surprising findings
- Euclidean-distance-based patch similarity performs worst in the PDL ablation, attributed to numerical instability.
- Adding patch-level distillation to the adapter-based model gives gains only around 1%, while the TSAI module alone gives most of the improvement over the linear-probe baseline.
Limitations
- Relies on a pre-trained ViT-B/16 and classifier alignment after each task, so gains depend on the PTM's generalizability.
- Gains from patch-level distillation and reconstruction are around 1% individually in ablations.
- Hyperparameters beta and lambda are tuned on CIFAR-100 and then used across datasets.
Experimental setup
Datasets: CIFAR-100 CUB-200 ImageNet-R ImageNet-A
Benchmarks: CIFAR-100 CUB-200 ImageNet-R ImageNet-A
Metrics: final accuracy A^T average incremental accuracy trainable parameters per task FLOPs per image
Baselines: L2P DualPrompt CODA-Prompt CPrompt ConvPrompt C-ADA LAE Adam-Adapter EASE InfLoRA SLCA Adam-Ft
Builds on: Adapter tuning SLCA EASE LAE CPrompt
Closest papers by meaning
- Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning (NeurIPS 2026 · similarity 0.89)
- Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning (CVPR 2025 · similarity 0.89)
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.87)
- Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.86)
- MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.