Research map / Architectures and efficiency / RNNs and normalization
Multigrid Neural Memory
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 multigrid memory instead of DNC | Mapping and localization · 25×25 maze, 3×3 FoV/query, spiral | F-score: 24.44 -> 99.17 | Helped |
| Restrict each layer to the finest pyramid scale (ablation) | Multigrid memory · Mapping and localization, 25×25 maze, spiral | F-score: 99.98 -> 61.73 | Hurt |
| Replace multigrid connectivity with stacked convolutional-LSTM layers | Mapping and localization · 25×25 maze, 3×3 FoV/query, spiral | F-score: 99.17 -> 6.51 | Hurt |
| Increase random-walk path length from 500 to 1500 | MG Mem+CNN · Mapping and localization, 25×25 maze | F-score: 96.20 -> 93.54 | Hurt |
| Use multigrid encoder-decoder instead of DNC | Priority Sort + Classify · MNIST, sequence length 20 | error rate: 0.1659±0.0188 -> 0.0864±0.0016 | Helped |
| Use multigrid memory instead of DNC | Associative Recall + Classify · MNIST, sequence length 10 | error rate: 0.2016±0.0161 -> 0.0736±0.0045 | Helped |
| Use multigrid memory instead of DNC | Priority Sort · Random vectors, sequence length 20 | error rate: 0.0039±0.0016 -> 0.0043±0.0009 | No clear effect |
| Use multigrid memory instead of DNC | Associative Recall · Random vectors, sequence length 10 | error rate: 0.0044±0.0001 -> 0.0030±0.0002 | Helped |
| Use multigrid memory instead of DNC | Question answering · bAbI, 20 tasks | mean error (%): 16.7 ± 7.6 -> 9.2±1.0 | Helped |
| Use multigrid memory instead of DNC | Question answering · bAbI, 20 tasks | failed tasks (error > 5%): 11.2 ± 5.4 -> 5.4±1.4 | Helped |
Closest papers by meaning
- Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks (ICLR 2024 · similarity 0.85)
- Blending Complementary Memory Systems in Hybrid Quadratic-Linear Transformers (NeurIPS 2025 · similarity 0.85)
- Memory Layers at Scale (ICML 2025 · similarity 0.84)
- Factual recall in linear associative memories: sharp asymptotics and mechanistic insights (NeurIPS 2026 · similarity 0.84)
- The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology (NeurIPS 2026 · similarity 0.84)
- Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons (NeurIPS 2026 · 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.