Research map / Continual learning / Online and replay-based
Online Learned Continual Compression with Adaptive Quantization Modules
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 ER with AQM memory | online continual classification · CIFAR-10 shared-head, M=20 | accuracy: 27.5± 1.2 -> 43.5 ± 0.7 | Helped |
| replace ER with AQM memory | online continual classification · CIFAR-10 shared-head, M=50 | accuracy: 33.1 ± 1.7 -> 47.0 ± 0.8 | Helped |
| use AQM instead of ER-MIR | online continual classification · CIFAR-10 shared-head, M=20 | forgetting: 50.2 ± 2.0 -> 23.0 ± 1.0 | Helped |
| use AQM instead of ER-MIR | online continual classification · CIFAR-10 shared-head, M=50 | forgetting: 30.2 ± 2.3 -> 19.0 ± 1.4 | Helped |
| add a second AQM module (ablation) | AQM · ImageNet offline training from online-compressed memory | accuracy: 20.5 ± 1.3 -> 23.2 ± 1.1 | Helped |
| fix the codebook after satisfactory compression is reached (ablation) | AQM · ImageNet offline training from online-compressed memory | accuracy: 19.2 ± 0.6 -> 23.2 ± 1.1 | Helped |
| use decoupled rather than end-to-end module training (ablation) | AQM · ImageNet offline training from online-compressed memory | accuracy: 16.5 ± 0.7 -> 23.2 ± 1.1 | Helped |
| use adaptive compression rather than always storing in the most compressed block (ablation) | AQM · ImageNet offline training from online-compressed memory | accuracy: 13.1 ± 3.2 -> 23.2 ± 1.1 | Helped |
| stabilize the codebook by freezing it (ablation) | VQ-VAE online compressor · mini-ImageNet stream, recon th=1 | Drift MSE: 1.01 ± 0.14 -> 0.49 ± 0.02 | Helped |
| stabilize the codebook by freezing it (ablation) | VQ-VAE online compressor · mini-ImageNet stream, recon th=2.5 | Drift MSE: 13.24 ± 5.36 -> 1.07 ± 0.05 | Helped |
| stabilize the codebook by freezing it (ablation) | VQ-VAE online compressor · mini-ImageNet stream, recon th=5 | Drift MSE: 55.31 ± 36.82 -> 1.69 ± 0.18 | Helped |
| stabilize the codebook by freezing it (ablation) | VQ-VAE online compressor · mini-ImageNet stream, recon th=2.5 | Streaming MSE: 0.63 ± 0.06 -> 0.81 ± 0.02 | Hurt |
Closest papers by meaning
- Towards Backward-Compatible Continual Learning of Image Compression (CVPR 2024 · similarity 0.89)
- Salient Frequency-aware Exemplar Compression for Resource-constrained Online Continual Learning (AAAI 2025 · similarity 0.87)
- Continual-MAE: Adaptive Distribution Masked Autoencoders for Continual Test-Time Adaptation (CVPR 2024 · similarity 0.86)
- Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling (ICLR 2025 · similarity 0.85)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 2024 · similarity 0.85)
- Adaptive Dataset Quantization (AAAI 2025 · 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.