Research map / Continual learning / Class-incremental
Navigating Semantic Drift in Task-Agnostic 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 MACIL, a task-agnostic class-incremental learning method on a frozen ViT-B/16 with task-specific LoRA that calibrates semantic drift via mean shift compensation, Mahalanobis-based covariance calibration, and patch-token self-distillation.
Problem
Class-incremental learning with pre-trained models must balance plasticity and stability without task IDs at inference; LoRA-based updates induce feature mean and covariance drift (semantic drift) that degrades old-class classification.
Approach
Estimates class mean shifts for all old classes from weighted embedding changes (weights by proximity to old class mean), aligns class-wise covariances via a Mahalanobis-distance loss between old and current networks, adds patch-token self-distillation, and retrains the classifier on Gaussian samples from calibrated statistics.
Key results
- On ImageNet-R (10 tasks), the method reaches 81.88% last accuracy and 85.95% average accuracy, versus 79.55% and 83.70% for the second-best SSIAT.
- On ImageNet-A, last accuracy is 64.14%, 1.49 points above SSIAT (62.65%).
- Ablation on ImageNet-R shows mean shift compensation plus covariance calibration gives about 2.3 points over the LoRA baseline (79.36 to 81.60 last accuracy), with patch distillation adding a further small gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add mean shift compensation (ablation) | Task-specific LoRA baseline with classifier alignment · ImageNet-R (10 tasks) | last accuracy: 79.36 ± 0.57 -> 80.81 ± 0.16 | Helped |
| add covariance calibration (ablation) | Task-specific LoRA baseline with classifier alignment · ImageNet-R (10 tasks) | last accuracy: 79.36 ± 0.57 -> 80.70 ± 0.32 | Helped |
| add mean shift compensation and covariance calibration (ablation) | Task-specific LoRA baseline with classifier alignment · ImageNet-R (10 tasks) | last accuracy: 79.36 ± 0.57 -> 81.60 ± 0.09 | Helped |
| add patch distillation to mean shift compensation and covariance calibration (ablation) | Task-specific LoRA with MSC and CC · ImageNet-R (10 tasks) | last accuracy: 81.60 ± 0.09 -> 81.88 ± 0.07 | Helped |
| use task-specific LoRA instead of task-shared LoRA (ablation) | LoRA structure · ImageNet-R | last accuracy: 81.15±0.19 -> 81.88±0.07 | Helped |
| use task-specific LoRA instead of hybrid Hydra-LoRA (ablation) | LoRA structure · ImageNet-R | last accuracy: 81.00±0.09 -> 81.88±0.07 | Helped |
| use task-specific LoRA instead of task-shared LoRA (ablation) | LoRA structure · CIFAR-100 | last accuracy: 91.53±0.19 -> 91.94±0.17 | Helped |
| use task-specific LoRA instead of hybrid Hydra-LoRA (ablation) | LoRA structure · CIFAR-100 | last accuracy: 91.48±0.13 -> 91.94±0.17 | Helped |
| use task-shared LoRA instead of task-specific LoRA (ablation) | LoRA structure · ImageNet-A | last accuracy: 64.14±0.58 -> 64.27±0.08 | Helped |
Surprising findings
- RanPAC, a first-session adaptation method, remains very strong on CIFAR-100 and CUB-200, which the authors attribute to overlap between the ImageNet pretraining data and these datasets.
- The task-specific, task-shared, and hybrid LoRA structures differ only marginally in performance.
Limitations
- Gains on CIFAR-100 and CUB-200 are marginal relative to other methods, and RanPAC is competitive there.
- Results depend on ViT-B/16 pretrained on ImageNet-21K, and the method is evaluated only on image classification benchmarks.
- Hyperparameters (distillation weight, LoRA rank, angular scale) were chosen by sensitivity analysis on the benchmarks.
Experimental setup
Datasets: ImageNet-R ImageNet-A CUB-200 CIFAR-100 ImageNet-21K
Benchmarks: ImageNet-R ImageNet-A CUB-200 CIFAR-100
Metrics: last accuracy average incremental accuracy
Baselines: L2P DualPrompt CODA-Prompt SLCA RanPAC OS-Prompt VQ-Prompt CPrompt EASE InfLoRA+CA SSIAT MOS
Builds on: LoRA SLCA InfLoRA HydraLoRA SSIAT
Closest papers by meaning
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.90)
- T-CIL: Temperature Scaling using Adversarial Perturbation for Calibration in Class-Incremental Learning (CVPR 2025 · similarity 0.88)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.88)
- Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning (NeurIPS 2025 · similarity 0.87)
- Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction for Exemplar-Free Class-Incremental Learning (ICML 2025 · similarity 0.87)
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.