Research map / Continual learning / Online and replay-based
Salient Frequency-aware Exemplar Compression for Resource-constrained Online Continual 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 SFEC, a JPEG-based exemplar compression framework for online class-incremental learning that quantizes less salient frequencies (from gradients of DCT coefficients) and evicts exemplars with large compression distortion, with negligible extra compute.
Problem
Exemplar compression for online class-incremental learning under limited compute and buffer memory; existing learned or optimization-based compression methods add forward/backward passes that delay learning from the stream, while plain JPEG introduces training-harmful artifacts.
Approach
Use gradient-based saliency over DCT spatial frequencies to set per-frequency JPEG quantization tables (quantizing less salient frequencies more aggressively), and perform compression-aware buffer eviction weighted by reconstruction error between raw and compressed images.
Key results
- In the fast streaming scenario SFEC improves over ER-ACE by 1.87% on CIFAR-100 and 2.61% on Mini-ImageNet (M=2000 region; e.g., 23.85 vs 21.74 on CIFAR-100 at M=2000).
- Plain JPEG reduces ER-ACE accuracy (e.g., 19.23 vs 21.74 on CIFAR-100 at M=2000), while SFEC increases it.
- Plugging SFEC into PCR improves final accuracy at all buffer sizes, e.g., 27.94 vs 26.85 on CIFAR-100 at M=2k.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ER-ACE with SFEC | ER-ACE · CIFAR-100 and Mini-ImageNet, fast streaming | final average accuracy: 1.87% and 2.61% | Helped |
| use JPEG compression without saliency-aware quantization or compression-aware buffer management (ablation) | ER-ACE · CIFAR-100, M=2000 | final average accuracy: 21.74 ± 0.4 -> 19.23 ± 1.0 | Hurt |
| use JPEG compression without saliency-aware quantization or compression-aware buffer management (ablation) | ER-ACE · Mini-ImageNet, M=2000 | final average accuracy: 17.41 ± 1.2 -> 14.18 ± 3.6 | Hurt |
| add saliency-aware frequency quantization to JPEG compression (ablation) | ER-ACE + JPEG · CIFAR-100, M=2000 | final average accuracy: 19.23 ± 1.0 -> 23.56 ± 0.4 | Helped |
| add saliency-aware frequency quantization to JPEG compression (ablation) | ER-ACE + JPEG · Mini-ImageNet, M=2000 | final average accuracy: 14.18 ± 3.6 -> 18.27 ± 3.0 | Helped |
| add compression-aware buffer management to JPEG + SFQ (ablation) | ER-ACE + JPEG + SFQ · CIFAR-100, M=2000 | final average accuracy: 23.56 ± 0.4 -> 23.85 ± 1.2 | Helped |
| add compression-aware buffer management to JPEG + SFQ (ablation) | ER-ACE + JPEG + SFQ · Mini-ImageNet, M=2000 | final average accuracy: 18.27 ± 3.0 -> 19.50 ± 1.8 | Helped |
| use SFEC instead of ER-ACE | ER-ACE · CIFAR-100, slow streaming, M=5000 | final average accuracy: 27.34 ± 0.7 -> 26.86 ± 0.9 | Mixed |
| use SFEC with PCR | PCR · CIFAR-100 and Mini-ImageNet, M=500–5000 | final average accuracy | Helped |
| vary λ in the saliency-aware quantization threshold (ablation) | SFEC · CIFAR-100, M=2000 | final average accuracy | No clear effect |
Surprising findings
- Existing exemplar compression methods (AQM, MRDC, CIM) often underperform the simple ER-ACE baseline in streaming settings because compression compute delays learning.
- MRDC lowers baseline accuracy at all buffer sizes even without delay in the slow streaming scenario, indicating JPEG artifacts hurt training despite more exemplars.
Limitations
- Compression artifacts still exist; saliency threshold lambda is set to 1 without tuning per dataset.
- Evaluated only on CIFAR-100 and Mini-ImageNet with a single backbone (Reduced ResNet18) and a single base OCIL method (ER-ACE) plus PCR.
Experimental setup
Datasets: CIFAR-100 Mini-ImageNet
Benchmarks: CIFAR-100 Mini-ImageNet
Metrics: final average accuracy
Baselines: ER-ACE AQM MRDC CIM MIR GSS JPEG PCR
Builds on: ER-ACE JPEG MRDC PCR Reservoir sampling
Closest papers by meaning
- Tensor Decomposition Based Memory-Efficient Incremental Learning (ICML 2025 · similarity 0.89)
- F-OAL: Forward-only Online Analytic Learning with Fast Training and Low Memory Footprint in Class Incremental Learning (NeurIPS 2024 · similarity 0.87)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 2024 · similarity 0.86)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.86)
- Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling (ICLR 2025 · similarity 0.86)
- DS-AL: A Dual-Stream Analytic Learning for Exemplar-Free Class-Incremental Learning (AAAI 2024 · 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.