Research map / Continual learning / Online and replay-based
Memory Aware Synapses: Learning what (not) to forget
Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, Tinne Tuytelaars
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 local Hebbian layerwise importance (l-MAS) instead of global gradients (ablation) | MAS · 2-task object recognition | average forgetting (%): 1% -> 3% | Hurt |
| compute gradients of squared l2 norm instead of vector output (ablation) | MAS · Flower -> Scenes and Birds | drop on first task (%): 0.51% -> 0.50% | No clear effect |
| compute importance on unlabeled test data instead of training data (ablation) | MAS · Flower -> Scenes | first-task accuracy (%): 77.24 -> 77.45 | No clear effect |
| compute importance on unlabeled test data instead of training data (ablation) | l-MAS · Flower -> Birds | first-task accuracy (%): 73.96 -> 74.48 | No clear effect |
| compute importance on both training and test data | MAS · 6DS fact learning | MAP on all facts: 0.29 -> 0.30 | Improved |
| compute importance only on the class of interest (adaptation) | MAS · Krishna Kamal flower, 8 tasks | forgetting on subset (%): 8% -> 2% | Improved |
| compute importance only on the class of interest (adaptation) | MAS · Krishna Kamal flower, 8 tasks | last task accuracy (%): 69% -> 84% | Improved |
| compute importance only on sport subset | MAS · 6DS sport subset | MAP at end of sequence: 0.20 -> 0.50 | Improved |
| compute importance on same subset used for evaluation | MAS · 6DS two-task adaptation | MAP on T11: 0.451 -> 0.472 | Improved |
| use MAS instead of SI | MAS · 8-task object recognition | average accuracy (%): 50.49 -> 52.695 | Improved |
| use MAS instead of SI | MAS · 6DS fact learning, 4 tasks | MAP on all facts: 0.25 -> 0.29 | Improved |
| use MAS instead of SI | MAS · Scenes -> Birds, first task | accuracy (%): 57.46 -> 57.61 | No clear effect |
Closest papers by meaning
- Lifelong Neural Predictive Coding: Learning Cumulatively Online without Forgetting (NeurIPS 2022 · similarity 0.90)
- Continual Learning Through Synaptic Intelligence (ICML 2017 · similarity 0.90)
- Memory by accident: a theory of learning as a byproduct of network stabilization (NeurIPS 2025 · similarity 0.90)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.89)
- Predictive Attractor Models (NeurIPS 2024 · similarity 0.88)
- Factual recall in linear associative memories: sharp asymptotics and mechanistic insights (NeurIPS 2026 · 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.
