Research map / Continual learning / Online and replay-based
Repairing Neural Networks by Leaving the Right Past Behind
Ryutaro Tanno, Melanie F. Pradier, Aditya Nori, Yingzhen Li
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 |
|---|---|---|---|
| use EWC-influence instead of SGD-influence for cause identification | EWC-Influence · CIFAR-10 label noise | computation time (s): 8301.1 -> 496.3 | Improved |
| use EWC-influence instead of SGD-influence | EWC-Influence · MNIST label noise | computation time (s): 185.1 -> 9.8 | Improved |
| use EWC-influence instead of linear influence (SA) | EWC-Influence · CIFAR-10 label noise | computation time (s): 1322.4 -> 496.3 | Improved |
| use EWC-influence instead of linear influence (GD) | EWC-Influence · MNIST adversarial poisoning | computation time (s): 10.0 -> 14.5 | No clear effect |
| use EWC-influence to rank causes vs other influence methods | EWC-Influence · MNIST/CIFAR-10 label noise | fraction of identified causes with wrong labels | Improved |
| remove top causes by EWC-influence then fine-tune vs other influence methods | EWC-Influence · MNIST/CIFAR-10 label noise | accuracy on failure set | Improved |
| remove many (about 10^3) identified causes vs remove only annotation-noise instances | all influence methods · MNIST label noise | failure accuracy | Improved |
| remove EWC-identified causes vs remove all injected random input noise | EWC-Influence · MNIST/CIFAR-10 random input noise | accuracy on failure set | Improved |
| remove identified causes vs randomly removing data | all influence methods · MNIST/CIFAR-10 label noise | accuracy on failure set | Improved |
| shrink the failure query set to 5% (about 20 examples) | all methods · MNIST label noise | failure accuracy | No clear effect |
| use EWC-deletion instead of Newton update removal | EWC-deletion · CIFAR-10 label noise | best failure accuracy: by 5% | Improved |
| use EWC-deletion instead of Newton update removal | EWC-deletion · CIFAR-10 label noise | best accuracy on remaining set: less than 1% difference | Mixed |
Closest papers by meaning
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.86)
- BECAME: Bayesian Continual Learning with Adaptive Model Merging (ICML 2025 · similarity 0.86)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.86)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 2025 · similarity 0.85)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.85)
- Where is the Truth? The Risk of Getting Confounded in a Continual World (ICML 2025 · similarity 0.84)
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.
