Research map / Continual learning / Class-incremental
FCS: Feature Calibration and Separation for Non-Exemplar 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 FCS for non-exemplar class incremental learning, combining an optimal-transport-based Feature Calibration Network that moves old-class prototypes into the new feature space with a Prototype-Involved Contrastive loss that separates new-class features from old prototypes.
Problem
Non-exemplar class incremental learning (NECIL) suffers catastrophic forgetting because stored old-class prototypes drift as the feature extractor changes, and new-class features overlap and distort old-class decision boundaries without exemplars.
Approach
A neural-network transport plan, trained with an optimal-transport cost between old- and new-model features on new data, calibrates stored prototypes into the new feature space; prototypes are then used as negatives in a supervised contrastive loss to push new classes away from old classes.
Key results
- FCS reaches 62.13% top-1 on CIFAR-100 with 5 stages, versus 58.12% for FeTrIL and 56.40% for PASS.
- At 20 stages on CIFAR-100, FCS achieves 58.36% against 52.48% for FeTrIL, a 4.69-point gain over the next best method.
- On ImageNet-Subset (10 stages), FCS reaches 61.76%, only 0.54 points above FeTrIL (61.22%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FCN to calibrate old-class prototypes (ablation) | FCS baseline · CIFAR-100, 5 stages | top-1 accuracy: 60.51 -> 61.77 | Helped |
| Add FCN to calibrate old-class prototypes (ablation) | FCS baseline · CIFAR-100, 10 stages | top-1 accuracy: 57.33 -> 59.73 | Helped |
| Add FCN to calibrate old-class prototypes (ablation) | FCS baseline · CIFAR-100, 20 stages | top-1 accuracy: 54.48 -> 58.31 | Helped |
| Add PIC to the baseline with FCN (ablation) | FCS · CIFAR-100, 5 stages | top-1 accuracy: 61.77 -> 62.13 | Helped |
| Add PIC to the baseline with FCN (ablation) | FCS · CIFAR-100, 10 stages | top-1 accuracy: 59.73 -> 60.39 | Helped |
| Add PIC to the baseline with FCN (ablation) | FCS · CIFAR-100, 20 stages | top-1 accuracy: 58.31 -> 58.36 | Helped |
| Use a single linear layer for FCN instead of a two-layer network (ablation) | FCS · CIFAR-100, 20 stages | top-1 accuracy: 56.06 -> 58.36; 56.57 -> 58.36 | Helped |
Surprising findings
- On ImageNet-Subset, the frozen-backbone FeTrIL is nearly on par with FCS, suggesting backbone freezing preserves knowledge well on large datasets.
- Simultaneously using old and new classification heads causes a clear accuracy drop relative to using them separately, which the authors attribute to old/new boundary intersection.
Limitations
- Gains on ImageNet-Subset over the frozen-backbone FeTrIL are marginal (0.54 points).
- Calibration network is trained only on current-stage data and the within-stage feature drift is only approximated.
- Average forgetting on CIFAR-100 is not best for FCS in some settings (e.g., 5 stages FCS 12.20 vs others; table shows lowest in most columns, but only one dataset reported).
Experimental setup
Datasets: CIFAR-100 TinyImageNet ImageNet-Subset
Benchmarks: CIFAR-100 TinyImageNet ImageNet-Subset
Metrics: top-1 accuracy average forgetting
Baselines: LwF PASS IL2A SSRE R-DFCIL EDG FeTrIL iCaRL EEIL
Builds on: PASS prototype augmentation supervised contrastive loss optimal transport FeTrIL
Closest papers by meaning
- Prospective Representation Learning for Non-Exemplar Class-Incremental Learning (NeurIPS 2024 · similarity 0.89)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.86)
- Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.86)
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.85)
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.