Research map / Architectures and efficiency / RNNs and normalization
Meta-Learning with Memory-Augmented Neural 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 |
|---|---|---|---|
| replace LSTM controller with feedforward controller (ablation) | MANN · Omniglot 5 classes string labels | accuracy at 2nd instance (%): 8.0 -> 69.5 | Helped |
| replace LSTM controller with feedforward controller (LRUA) (ablation) | MANN · Omniglot 15 classes string labels | accuracy at 2nd instance (%): 12.8 -> 62.6 | Helped |
| use LRU Access instead of NTM access module (ablation) | MANN with LSTM controller · Omniglot 15 classes string labels | accuracy at 2nd instance (%): 35.4 -> 62.6 | Helped |
| use LRU Access instead of NTM access module (ablation) | MANN with LSTM controller · Omniglot 15 classes string labels | accuracy at 10th instance (%): 88.4 -> 95.3 | Helped |
| add external memory to LSTM (ablation) | LSTM · Omniglot 5 classes one-hot labels | accuracy at 2nd instance (%): 49.5 -> 82.8 | Helped |
| MANN vs human | MANN · Omniglot 5 classes one-hot labels | accuracy at 5th instance (%): 81.4 -> 94.9 | Helped |
| MANN vs kNN with autoencoder features | MANN · Omniglot 5 classes string labels | accuracy at 2nd instance (%): 51.9 -> 69.5 | Helped |
| MANN vs kNN with autoencoder features | MANN · Omniglot 15 classes string labels | accuracy at 10th instance (%): 60.0 -> 95.3 | Helped |
| use autoencoder deep features instead of raw pixels for kNN (ablation) | kNN · Omniglot 5 classes string labels | accuracy at 2nd instance (%): 36.7 -> 51.9 | Helped |
| do not wipe external memory between episodes (ablation) | MANN · Omniglot 5 classes one-hot labels | accuracy at 2nd instance (%) | Hurt |
| do not wipe external memory between episodes (ablation) | MANN · Omniglot 10 classes episode length 75 | accuracy at 2nd instance (%) | No clear effect |
| curriculum training raising max classes each 10,000 episodes (ablation) | MANN · Omniglot up to 100 classes | accuracy at 2nd instance (%) | Mixed |
Closest papers by meaning
- Learning Universal Predictors (ICML 2024 · similarity 0.89)
- Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes (NeurIPS 2016 · similarity 0.88)
- Metalearned Neural Memory (NeurIPS 2019 · similarity 0.88)
- Neural Stored-program Memory (ICLR 2020 · similarity 0.88)
- Blending Complementary Memory Systems in Hybrid Quadratic-Linear Transformers (NeurIPS 2025 · similarity 0.86)
- Memory-Reduced Meta-Learning with Guaranteed Convergence (AAAI 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.