Research map / Continual learning / Online and replay-based
Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?
Mohammad Ali Vahedifar, Abhisek Ray, Qi Zhang
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 |
|---|---|---|---|
| select neurons by Shapley value instead of binary WSN selection (ablation) | SNV · CIFAR-100 TIL (10 tasks), c=0.5 | ACC: 64.00 -> 79.76 | Improved |
| select neurons by Shapley value instead of WSN selection (ablation) | SNV · Tiny-ImageNet TIL (10 tasks), c=0.5 | ACC: 61.06 -> 74.82 | Improved |
| Shapley selection vs WSN under tight budget c=0.03 (ablation) | SNV · CIFAR-100 TIL (10 tasks) | ACC: 59.65 -> 71.74 | Improved |
| increase per-task capacity ratio c from 0.03 to 0.5 (ablation) | SNV · CIFAR-100 TIL (10 tasks) | ACC: 71.74 -> 79.76 | Improved |
| increase capacity c from 0.1 to 0.3 on Tiny-ImageNet (ablation) | SNV · Tiny-ImageNet TIL (10 tasks) | ACC: 74.73 -> 74.38 | No clear effect |
| mask-based freezing of Shapley-selected neurons | ResNet-18 · ImageNet-1k CIL (10 tasks) | BWT: -37.78 (NFL+) -> -0.05 | Improved |
| Shapley neuron freezing instead of NFL+ progressive parameter freezing | ResNet-18 · ImageNet-1k CIL (10 tasks) | ACC: 38.42 -> 41.30 | Improved |
| SNV vs memory-based DyTox with 20000 exemplars | ResNet-18 · ImageNet-1k TIL (10 tasks) | ACC: 59.40 (DyTox) vs 57.82 | Mixed |
| add Shapley estimation step (Monte Carlo, truncation, bandit) | NFL+ baseline · ImageNet-1k CIL (10 tasks) compute cost | FLOPs: 1,221 PF -> 1,514 PF; 38 h -> 47 h | Hurt |
| add Shapley estimation (SNV) over NFL+ | ResNet-18 · CIFAR-100 CIL (10 tasks) | ACC: 53.70 -> 54.70 | Improved |
Closest papers by meaning
- Mitigating Forgetting in Online Continual Learning with Neuron Calibration (NeurIPS 2021 · similarity 0.87)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.85)
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.85)
- Self-Normalized Resets for Plasticity in Continual Learning (ICLR 2025 · similarity 0.85)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.84)
- Optimal Protocols for Continual Learning via Statistical Physics and Control Theory (ICLR 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.
