Research map / Architectures and efficiency / Pruning and efficient tuning
Learning from Oblivion: Predicting Knowledge-Overflowed Weights via Retrodiction of Forgetting
Jinhyeok Jang, Jaehong Kim, Jung Uk Kim
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 KNOW prediction with KNOWN (x8) | ResNet18 CIFAR100 pretrain · CIFAR10 (100% pretrain data) | accuracy: 92.40 -> 93.55 | Improved |
| add KNOW prediction with KNOWN (x8) | ResNet18 CIFAR100 pretrain · CIFAR10 (6.25% pretrain data) | accuracy: 91.51 -> 92.07 | Improved |
| incrementally pretrain 6.25% to 100% instead of naive transfer | ResNet18 · CIFAR10 (6.25% data) | accuracy: 91.51 -> 89.51 | Hurt |
| use exponential curve fit for weight extrapolation (ablation) | KNOW · CIFAR10 (x8) | accuracy: 92.40 -> 66.54 | Hurt |
| use Task Vector extrapolation instead of KNOWN (ablation) | KNOW · CIFAR10 (x8, 100% data) | accuracy: 93.55 -> 92.65 | Hurt |
| increase history length S from 2 to 5 (ablation) | KNOWN x8 · CIFAR10 | accuracy: 92.65 -> 93.55 | Improved |
| add KNOWN x9 | PVTv2 · TinyImageNet | accuracy: 76.17 -> 77.58 | Improved |
| add KNOWN x3 | PVTv2 · CIFAR100 | accuracy: 82.03 -> 82.46 | No clear effect |
| add TSV x3 and x9 (ablation) | PVTv2 · Flowers | accuracy: 87.98 -> 86.37 (x3), 86.47 (x9) | Hurt |
| add KNOWN | PVTv2 · PACS leave-one-domain-out average | accuracy: 63.48 -> 68.33 (x9) | Improved |
| add MagMax x9 or TSV x9 (ablation) | PVTv2 · PACS average | accuracy: 63.48 -> 63.17 (MagMax); 63.45 (TSV) | No clear effect |
| add KNOWN x3 | PVTv2 · Flickr8K captioning | masked accuracy: 37.14 -> 39.38 | Improved |
Closest papers by meaning
- Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting (ICML 2025 · similarity 0.86)
- On Local Overfitting and Forgetting in Deep Neural Networks (AAAI 2025 · similarity 0.86)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.86)
- Reinforced Fast Weights via Next-Sequence Prediction (NeurIPS 2026 · similarity 0.86)
- Knowledge Swapping via Learning and Unlearning (ICML 2025 · similarity 0.85)
- LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning (ICML 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.
