Research map / Continual learning / Online and replay-based
Improved Schemes for Episodic Memory-based Lifelong Learning
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 A-GEM's fixed emphasis with adaptive loss balancing and gradient mixing (MEGA-1) | episodic-memory lifelong learning · Permuted MNIST | average accuracy: 89.32 $\pm$ 0.46 -> 91.10 $\pm$ 0.08 | Helped |
| replace A-GEM's fixed emphasis with adaptive loss balancing and gradient mixing (MEGA-1) | episodic-memory lifelong learning · Split CIFAR | average accuracy: 61.28 $\pm$ 1.88 -> 66.10 $\pm$ 1.67 | Helped |
| replace A-GEM's fixed emphasis with adaptive loss balancing and gradient mixing (MEGA-1) | episodic-memory lifelong learning · Split CUB | average accuracy: 61.82 $\pm$ 3.72 -> 79.67 $\pm$ 2.15 | Helped |
| replace A-GEM's fixed emphasis with adaptive loss balancing and gradient mixing (MEGA-1) | episodic-memory lifelong learning · Split AWA | average accuracy: 44.95 $\pm$ 2.97 -> 54.82 $\pm$ 4.97 | Helped |
| use adaptive loss balancing rather than fixed α1=1 and α2=1 (ablation) | MEGA-1 · Permuted MNIST, Split CIFAR, Split CUB, Split AWA | average accuracy: 90.66 $\pm$ 0.09 -> 91.10 $\pm$ 0.08; 64.65 $\pm$ 1.98 -> 66.10 $\pm$ 1.67; 79.44 $\pm$ 2.98 -> 79.67 $\pm$ 2.15; 53.60 | Helped |
| use loss-aware gradient rotation rather than equal current and reference losses (ablation) | MEGA-2 · Permuted MNIST, Split CIFAR, Split CUB, Split AWA | average accuracy: 91.15 $\pm$ 0.12 -> 91.21 $\pm$ 0.10; 58.04 $\pm$ 1.89 -> 66.12 $\pm$ 1.93; 68.60 $\pm$ 1.98 -> 80.58 $\pm$ 1.94; 47.95 | Helped |
| use MEGA-2 gradient rotation with equal current and reference losses instead of A-GEM's update (ablation) | episodic-memory lifelong learning · Split CIFAR | average accuracy: 61.28 $\pm$ 1.88 -> 58.04 $\pm$ 1.89 | Hurt |
| use MEGA-2 gradient rotation with equal current and reference losses instead of A-GEM's update (ablation) | episodic-memory lifelong learning · Permuted MNIST, Split CUB, Split AWA | average accuracy: 89.32$\pm$0.46 -> 91.15 $\pm$ 0.12; 61.82$\pm$3.72 -> 68.60 $\pm$1.98; 44.95$\pm$2.97 -> 47.95 $\pm$4.54 | Helped |
| use MEGA-2 rather than MEGA-1 when examples per task are limited | episodic-memory lifelong learning · Permuted MNIST tasks with limited examples | average accuracy | Helped |
| use MEGA-2 rather than MEGA-1 across benchmarks | episodic-memory lifelong learning · Permuted MNIST, Split CIFAR, Split CUB, Split AWA | average accuracy: 91.10 $\pm$ 0.08 vs 91.21 $\pm$ 0.10; 66.10 $\pm$ 1.67 vs 66.12 $\pm$ 1.93; 79.67 $\pm$ 2.15 vs 80.58 $\pm$ 1.94; 54.82 | Mixed |
Closest papers by meaning
- Hyper-LifelongGAN: Scalable Lifelong Learning for Image Conditioned Generation (CVPR 2021 · similarity 0.89)
- Gradient Episodic Memory for Continual Learning (NeurIPS 2017 · similarity 0.88)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.86)
- Adaptive Retention & Correction: Test-Time Training for Continual Learning (ICLR 2025 · similarity 0.85)
- Privacy-Aware Lifelong Learning (ICLR 2025 · similarity 0.85)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · 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.