Research map / Continual learning / Class-incremental
Multi-layer Rehearsal 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 MRFA, which augments rehearsal-sample features at every block via a gradient-ascent step to enlarge the all-layer margin and reduce forgetting in class-incremental learning.
Problem
Rehearsal-based class-incremental learning overfits to replay samples, and their all-layer margin shrinks as tasks progress, hurting generalization on old classes.
Approach
Augment each rehearsal sample's intermediate feature at a randomly chosen block input by a gradient-ascent step (scaled by random L2 norm of the feature), so training on augmented features enlarges the all-layer margin of rehearsal samples; plug-in to existing rehearsal/CIL methods.
Key results
- MRFA improves Replay, iCaRL, FOSTER and DyTox+ on CIFAR100 and ImageNet100 in most settings, typically by about 1-2%.
- Example: iCaRL on CIFAR100 10-10 with 2000 memory goes from 64.52 to 65.17 average accuracy with MRFA.
- Augmenting only rehearsal samples works better than augmenting both current and rehearsal samples (MRFA-a), e.g. 51.61 vs 50.16 for Replay at 500 memory, 10-10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MRFA to rehearsal-based CIL methods | CIFAR100 and ImageNet100 baselines · CIFAR100 and ImageNet100, listed memory sizes and task splits | average accuracy | Helped |
| use random selection and eviction instead of herding for rehearsal samples (ablation) | Replay with MRFA · CIFAR100 10-10, memory 500 | average accuracy: 54.20 -> 55.48 | Helped |
| use random selection and eviction instead of herding for rehearsal samples (ablation) | Replay with MRFA · CIFAR100 50-10, memory 500 | average accuracy: 46.61 -> 47.86 | Helped |
| augment current-task samples as well as rehearsal samples (ablation) | Replay with MRFA · CIFAR100 10-10, memory 500 | average accuracy: 51.61 -> 50.16 | Hurt |
| augment current-task samples as well as rehearsal samples (ablation) | Replay with MRFA · CIFAR100 50-10, memory 500 | average accuracy: 42.85 -> 41.03 | Hurt |
| augment current-task samples as well as rehearsal samples (ablation) | iCaRL with MRFA · CIFAR100 10-10, memory 500 | average accuracy: 54.84 -> 52.31 | Hurt |
| augment current-task samples as well as rehearsal samples (ablation) | iCaRL with MRFA · CIFAR100 50-10, memory 500 | average accuracy: 51.48 -> 49.85 | Hurt |
| perform the MRFA augmentation pass every p batches rather than every batch (ablation) | Replay and iCaRL with MRFA · CIFAR100 50-10 | average incremental accuracy | No clear effect |
| vary the maximum augmentation scale β (ablation) | Replay and iCaRL with MRFA · CIFAR100 10-10 and 50-10 | average accuracy | Mixed |
Surprising findings
- Applying the augmentation to current-task samples as well (MRFA-a) degrades performance, attributed to imbalance between rehearsal and current-task samples.
Limitations
- Each batch requires an extra forward-backward pass, adding training cost (mitigated by periodic application).
- Hyperparameter beta must be tuned per setting (best around 1e-4 for 10-10 and 1e-3 for 50-10).
Experimental setup
Datasets: CIFAR100 ImageNet100
Benchmarks: CIFAR100 ImageNet100
Metrics: last-task accuracy average incremental accuracy
Baselines: Replay iCaRL FOSTER DyTox+
Builds on: iCaRL FOSTER DyTox all-layer margin
Closest papers by meaning
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.87)
- Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective (ICML 2025 · similarity 0.87)
- Make Domain Shift a Catastrophic Forgetting Alleviator in Class-Incremental Learning (AAAI 2025 · similarity 0.87)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.86)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.86)
- Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 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.