Research map / Continual learning / Online and replay-based
NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks
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 |
|---|---|---|---|
| Use type-2-only Transfer growth rather than random growth (ablation) | NISPA · Permuted MNIST, 10% permutation | average accuracy across all tasks: 97.2 -> 97.3 | No clear effect |
| Use type-2-only Transfer growth rather than random growth (ablation) | NISPA · Permuted MNIST, 100% permutation | average accuracy across all tasks: 97.2 -> 97.0 | Hurt |
| Use type-1-only Novel growth rather than random growth (ablation) | NISPA · Permuted MNIST, 10% permutation | average accuracy across all tasks: 97.2 -> 97.3 | No clear effect |
| Use type-1-only Novel growth rather than random growth (ablation) | NISPA · Permuted MNIST, 100% permutation | average accuracy across all tasks: 97.2 -> 97.3 | No clear effect |
| Use cosine annealing rather than linear τ decrease while retaining re-initialization (ablation) | NISPA · EF-MNIST | average accuracy across five runs and all tasks: 91.0±0.1 -> 91.6±0.2 | Helped |
| Re-initialize weights at task boundaries while using cosine annealing (ablation) | NISPA · EF-MNIST | average accuracy across five runs and all tasks: 90.7±0.3 -> 91.6±0.2 | Helped |
| Use cosine annealing rather than linear τ decrease while retaining re-initialization (ablation) | NISPA · CIFAR100 | average accuracy across five runs and all tasks: 74.7±1.0 -> 75.9±1.3 | Helped |
| Re-initialize weights at task boundaries while using cosine annealing (ablation) | NISPA · CIFAR100 | average accuracy across five runs and all tasks: 74.9±0.6 -> 75.9±1.3 | Helped |
| Use NISPA-Replay rather than the strongest reported replay baseline | Class-incremental learning · MNIST, 10 samples per class | average accuracy across tasks: 73±1.4 -> 77±1.2 | Helped |
| Use DER rather than NISPA-Replay | Class-incremental learning · EF-MNIST, 50 samples per class | average accuracy across tasks: 74±0.7 -> 77±0.3 | Helped |
| Use iCaRL rather than NISPA-Replay | Class-incremental learning · CIFAR10, 10 samples per class | average accuracy across tasks: 38 ±2.0 -> 40±2.7 | Helped |
| Use NISPA-Replay rather than the strongest reported replay baseline | Class-incremental learning · CIFAR10, 100 samples per class | average accuracy across tasks: 30±1.8 -> 57±0.8 | Helped |
Closest papers by meaning
- Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective (ICML 2025 · similarity 0.90)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Self-Normalized Resets for Plasticity in Continual Learning (ICLR 2025 · similarity 0.87)
- InTAct: Interval-based Task Activation Consolidation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning (CVPR 2025 · similarity 0.87)
- Recall-Oriented Continual Learning with Generative Adversarial Meta-Model (AAAI 2024 · similarity 0.87)
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.