Research map / Continual learning / Class-incremental
Cross-Class Feature Augmentation 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 Cross-Class Feature Augmentation (CCFA), which uses adversarial-attack-style PGD perturbations of current features toward old classes, pseudo-labeled by the previous classifier, to augment class incremental learning and plugs into UCIR, PODNet, AANet and AFC.
Problem
Catastrophic forgetting in class incremental learning, where models must learn new classes without forgetting old ones with limited exemplar memory.
Approach
Augment features of arbitrary target classes by running a PGD attack on current features using the previous-stage classifier toward a chosen old class, then assign pseudo-labels from the old classifier, thereby using previous knowledge to complement training examples across classes.
Key results
- On CIFAR-100 with 50 tasks, UCIR+CCFA improves average incremental accuracy from 50.44 to 52.64, and PODNet+CCFA from 57.84 to 60.69.
- CCFA consistently improves all four baseline methods (UCIR, PODNet, AANet, AFC) across CIFAR-100 task sizes 50, 25, 10 and 5 in the reported table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CCFA to UCIR* | UCIR* · CIFAR-100, 50/25/10/5 tasks | average incremental accuracy: 50.44 ±0.32 -> 52.64 ±1.17; 56.61 ±0.32 -> 57.42 ±0.70; 60.47 ±0.32 -> 61.58 ±0.60; 63.76 ±0.32 -> 64.13±0.32 | Helped |
| Add CCFA to PODNet* | PODNet* · CIFAR-100, 50/25/10/5 tasks | average incremental accuracy: 57.84 ±0.37 -> 60.69 ±0.56; 60.50 ±0.62 -> 62.91 ±0.91; 62.77 ±0.78 -> 65.50 ±0.70; 64.62 ±0.65 -> 67.24 ±0.70 | Helped |
| Add CCFA to AANet* | AANet* · CIFAR-100, 50/25/10/5 tasks | average incremental accuracy: 60.91±1.00 -> 62.20 ±1.31; 62.34 ±1.17 -> 63.74 ±0.92; 64.49 ±0.81 -> 66.16±1.03; 66.34 ± 0.76 -> 67.37 ± 1.12 | Helped |
| Add CCFA to AFC* | AFC* · CIFAR-100, 50/25/10/5 tasks | average incremental accuracy: 61.74 ±0.68 -> 63.11 ±0.18; 63.78 ±0.42 -> 64.59±0.41; 64.63 ±0.51 -> 65.61 ±0.57; 66.02 ±0.47 -> 66.47 ±0.62 | Helped |
Limitations
- Target selection is an approximation; full target optimization is not used.
- Number of augmented features must be tuned for stability-plasticity trade-off.
Experimental setup
Datasets: CIFAR-100 ImageNet-100 ImageNet-1000
Benchmarks: CIFAR-100 ImageNet-100 ImageNet-1000
Metrics: average incremental accuracy accuracy per incremental stage
Baselines: UCIR PODNet AANet AFC
Builds on: UCIR PODNet AANet AFC
Closest papers by meaning
- Identifying and Understanding Cross-Class Features in Adversarial Training (ICML 2025 · similarity 0.87)
- Enhancing Robustness in Incremental Learning with Adversarial Training (AAAI 2025 · similarity 0.86)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.86)
- Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning (ICML 2024 · similarity 0.85)
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.85)
- Task Confusion and Catastrophic Forgetting in Class-Incremental Learning: A Mathematical Framework for Discriminative and Generative Modelings (NeurIPS 2024 · similarity 0.84)
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.