Research map / Architectures and efficiency / RNNs and normalization
Resurrecting Recurrent Neural Networks for Long Sequences
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 |
|---|---|---|---|
| remove recurrent nonlinearity (use linear RNN recurrence) (ablation) | vanilla dense RNN in deep architecture · sCIFAR, ListOps, Text, Retrieval | test accuracy: sCIFAR +2.3 (69.9 to 72.2); ListOps +6.5 (43.9 to 50.4) | Helped |
| diagonalize linear RNN into complex diagonal form with uniform disk initialization (ablation) | dense linear RNN · sCIFAR, ListOps | test accuracy: sCIFAR 72.2 to 86.5; ListOps 50.4 to 58.8 | Helped |
| exponential parameterization of eigenvalue magnitude and phase (no stability enforcement) (ablation) | diagonal complex RNN · ListOps | test accuracy: 58.8 to 60.5 | Helped |
| exponential parameterization without stability enforcement (ablation) | diagonal complex RNN · sCIFAR | test accuracy: 86.5 to 85.4 | Hurt |
| enforce stability via positive nonlinearity in exponential parameterization (ablation) | diagonal complex RNN with exp param · Pathfinder | test accuracy: 65.4 to 93.5 | Helped |
| enforce stability via positive nonlinearity in exponential parameterization (ablation) | diagonal complex RNN with exp param · sCIFAR | test accuracy: 85.4 to 87.2 | Helped |
| ring initialization of eigenvalues closer to unit circle (tuned r_min, r_max) (ablation) | stable exponential diagonal RNN · Pathfinder | test accuracy: 93.5 to 94.4 | Helped |
| ring initialization of eigenvalues closer to unit circle (tuned r_min, r_max) (ablation) | stable exponential diagonal RNN · sCIFAR | test accuracy: 87.2 to 88.1 | Helped |
| ring initialization of eigenvalues closer to unit circle (tuned r_min, r_max) (ablation) | stable exponential diagonal RNN · ListOps | test accuracy: 59.4 unchanged | No clear effect |
| add gamma normalization of input projection (ablation) | ring-initialized stable diagonal RNN (LRU) · PathX | test accuracy: unable to learn to random chance without normalization to 94.2 | Helped |
| add gamma normalization of input projection (ablation) | ring-initialized stable diagonal RNN (LRU) · sCIFAR | test accuracy: 88.1 to 89.0 | Helped |
| reduce eigenvalue phase initialization range to [0, pi/10] (ablation) | LRU on PathX · PathX | test accuracy: to 94.2 from random chance | Helped |
Closest papers by meaning
- Parallelizing Linear Recurrent Neural Nets Over Sequence Length (ICLR 2018 · similarity 0.91)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.90)
- Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvalues (ICML 2024 · similarity 0.89)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.88)
- Improving Bilinear RNN with Closed-loop Control (NeurIPS 2025 · similarity 0.88)
- Longhorn: State Space Models are Amortized Online Learners (ICLR 2025 · similarity 0.88)
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.