Research map / Continual learning / Class-incremental
FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning
Dipam Goswami, Yuyang Liu, Bartłomiej Twardowski, Joost van de Weijer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Euclidean NCM with per-class Mahalanobis (FeCAM) | Eucl-NCM · CIFAR-100 T=5 | avg accuracy (%): 64.8 -> 70.9 | Improved |
| use per-class covariance instead of a common covariance (ablation) | FeCAM · TinyImageNet T=5 | avg accuracy (%): 56.0 -> 59.6 | Improved |
| use FeCAM instead of FeTrIL | FeTrIL · ImageNet-Subset T=5 | avg accuracy (%): 73.1 -> 78.3 | Improved |
| use FeCAM instead of L2P and NCM with pretrained ViT | L2P / NCM · CoRe50 | accuracy (%): 78.3 -> 89.9 | Improved |
| use FeCAM without exemplars vs exemplar-based DER | DER · CIFAR-100 T=5 | last accuracy (%): 66.2 -> 62.1 | Hurt |
| add Tukey transformation to Euclidean NCM (ablation) | Euclidean NCM · CIFAR-100 T=5 | avg accuracy (%): 64.8 -> 66.6 | Improved |
| add covariance shrinkage to full Mahalanobis (ablation) | Mahalanobis · CIFAR-100 T=5 | avg accuracy (%): 29.7 -> 59.3 | Improved |
| add Tukey transformation to full Mahalanobis with shrinkage (ablation) | Mahalanobis · CIFAR-100 T=5 | avg accuracy (%): 59.3 -> 62.8 | Improved |
| add correlation normalization (ablation) | Mahalanobis with Tukey and shrinkage · CIFAR-100 T=5 | avg accuracy (%): 62.8 -> 70.9 | Improved |
| use full covariance instead of diagonal (ablation) | FeCAM · CIFAR-100 T=5 | avg accuracy (%): 66.9 -> 70.9 | Improved |
| use Bayes Mahalanobis classifier instead of sampling features and training a linear classifier (ablation) | FeCAM · CIFAR-100 T=5 | avg accuracy | Improved |
| use FeCAM instead of FeTrIL with small first task (20 classes) | FeTrIL · ImageNet-Subset | avg accuracy (%): 63.1 -> 66.4 | Improved |
Later papers that compared with this method
7 later accepted papers report FeCAM as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.
- EFC++: Elastic Feature Consolidation with Prototype Re-balancing for Cold Start Exemplar-free Incremental Learning (IJCV 2026 · better on 4 of 4 shared results)
- HyCal: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning (CVPR 2026 · better on 2 of 2 shared results)
- PANDA – Patch and Distribution-Aware Augmentation for Long-Tailed Exemplar-Free Continual Learning (AAAI 2026 · better on 2 of 2 shared results)
- Continual Out-of-Distribution Detection with Analytic Neural Collapse (AAAI 2026 · better on 1 of 1 shared result)
- Adapt Before Continual Learning (AAAI 2026 · better on 1 of 1 shared result)
- Spectral-Aware Analytic Class-Incremental Learning for Long-Tailed Distributions (ECCV 2026 · better on 1 of 1 shared result)
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR 2024 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.90)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.89)
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.89)
- Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning (ICLR 2024 · similarity 0.89)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.88)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.88)
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.
