Research map / Continual learning / Online and replay-based
Is Forgetting Less a Good Inductive Bias for Forward Transfer?
Jiefeng Chen, Timothy Nguyen, Dilan Gorur, Arslan Chaudhry
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 |
|---|---|---|---|
| Retain more knowledge from previous tasks during continual training | Continual learning representations · Split CIFAR-100, random init | K-shot forward transfer: 0.92 (Spearman correlation, k=5) | Improved |
| Retain more knowledge from previous tasks during continual training | Continual learning representations · Split CIFAR-10, random init | K-shot forward transfer: 0.64 (k=5); 0.56 (k=10); 0.65 (k=20) | Improved |
| Retain more knowledge from previous tasks during continual training | Continual learning representations · CLEAR100, pretrained | K-shot forward transfer: 0.83 (k=5); 0.8 (k=10); 0.81 (k=20) | Improved |
| Retain more knowledge from previous tasks during continual training | Continual learning representations · CIFAR-100 Superclasses, pretrained, k=5 | K-shot forward transfer: 0.16 (0.40) | No clear effect |
| Use LP-FT, which adds linear probing before finetuning each task | ImageNet-pretrained representations · Split CIFAR-10 and Split CIFAR-100 | K-shot forward transfer | Mixed |
| Use ResNet18 instead of ResNet50 (ablation) | Continual learning representations · Split CIFAR-10, Split CIFAR-100, and CIFAR-100 Superclasses | Relationship between forgetting and forward transfer | No clear effect |
| Measure transfer with auxiliary k-shot evaluation on fixed representations rather than accuracy after constrained continual updates | Forward-transfer evaluation · Image classification benchmarks | Association between knowledge retention and forward transfer: Most entries are above 0.5 | Improved |
| Retain less information from previous tasks | Continual learning representations · CIFAR-100 Superclasses, pretrained, k=5 | K-shot forward transfer: 0.16 (0.40) | No clear effect |
Closest papers by meaning
- Probing Representation Forgetting in Supervised and Unsupervised Continual Learning (CVPR 2022 · similarity 0.92)
- Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective (ICML 2025 · similarity 0.91)
- Disentangling and mitigating the impact of task similarity for continual learning (NeurIPS 2024 · similarity 0.90)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.89)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.88)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · 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.
