Research map / Continual learning / Class-incremental
Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learning
James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, Zsolt Kira
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 |
|---|---|---|---|
| add task balancing loss weighting (ablation) | Ours · CIFAR-100 (10 tasks, ablation) | final accuracy A_N (%): 23.4 -> 33.7 | Improved |
| use modified cross-entropy loss instead of standard CE (ablation) | Ours · CIFAR-100 (10 tasks, ablation) | final accuracy A_N (%): 16.5 -> 33.7 | Improved |
| add real data distillation (ablation) | Ours · CIFAR-100 (10 tasks, ablation) | final accuracy A_N (%): 15.9 -> 33.7 | Improved |
| add synthetic data distillation (ablation) | Ours · CIFAR-100 (10 tasks, ablation) | final accuracy A_N (%): 12.7 -> 33.7 | Improved |
| add fine-tuning cross-entropy (FT-CE) loss (ablation) | Ours · CIFAR-100 (10 tasks, ablation) | final accuracy A_N (%): 9.8 -> 33.7 | Improved |
| use the proposed method instead of DeepInversion distillation | Ours · CIFAR-100 (10 tasks, data-free) | final accuracy A_N (%): 10.9 -> 33.7 | Improved |
| add synthetic data to LwF | LwF · CIFAR-100 (10 tasks, data-free) | final accuracy A_N (%): 9.2 -> 8.9 | No clear effect |
| add generative replay (DGR) | Base · CIFAR-100 (10 tasks, data-free) | final accuracy A_N (%): 8.8 -> 8.1 | Hurt |
| use the proposed data-free method instead of LwF with coreset replay | Ours · CIFAR-100 (10 tasks, with replay coreset) | final accuracy A_N (%): 27.4 -> 33.7 | Improved |
| store coreset replay and use BiC instead of the data-free method | BiC · CIFAR-100 (10 tasks, with replay coreset) | final accuracy A_N (%): 33.7 -> 45.9 | Improved |
| proposed method vs LwF.MC at large resolution | Ours · ImageNet-50 (5 tasks) | final accuracy A_N (%): 72.7 -> 71.5 | Hurt |
| use the proposed method instead of LwF without replay | Ours · ImageNet-50 (5 tasks) | final accuracy A_N (%): 19.4 -> 71.5 | Improved |
Closest papers by meaning
- Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (TPAMI 2022 · similarity 0.91)
- IL2M: Class Incremental Learning With Dual Memory (ICCV 2019 · similarity 0.91)
- Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · similarity 0.91)
- Few-Shot Class-Incremental Learning via Class-Aware Bilateral Distillation (CVPR 2023 · similarity 0.91)
- SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection (CVPR 2024 · similarity 0.88)
- Tensor Decomposition Based Memory-Efficient Incremental Learning (ICML 2025 · 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.
