Research map / Architectures and efficiency / RNNs and normalization
Lattice Recurrent Unit: Improving Convergence and Statistical Efficiency for Sequence Modeling
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 |
|---|---|---|---|
| Decouple projected states, reset gates, and update gates to form LRU | GRU family · PTB, WP, enwik8, and text8 | test CCE: 0.972 -> 0.965; 1.163 -> 1.141; 1.101 -> 1.057; 1.047 -> 1.010 | Helped |
| Decouple projected states, reset gates, and update gates to form LRU | GRU family · PTB and WP (10M parameters) | epochs to convergence: 9 -> 8; 11 -> 8 | Helped |
| Decouple projected states, reset gates, and update gates to form LRU | GRU family · enwik8 and text8 (10M parameters) | epochs to convergence: 37 -> 14; 17 -> 17 | Mixed |
| Decouple projected states, reset gates, and update gates to form LRU | GRU family · text8 (10M parameters) | test CCE: 1.009 -> 1.010 | Hurt |
| Decouple projected states while keeping reset and update gates shared (ablation) | LRU · PTB, WP, enwik8, and text8 (10M parameters) | test CCE: 0.965 -> 0.971; 1.141 -> 1.148; 1.057 -> 1.090; 1.010 -> 1.019 | Hurt |
| Decouple reset gates and projected states while keeping update gates shared (ablation) | LRU · PTB, WP, enwik8, and text8 (10M parameters) | test CCE: 0.965 -> 0.971; 1.141 -> 1.143; 1.057 -> 1.074; 1.010 -> 1.014 | Hurt |
| Use separate output flows for time and depth rather than a single GRU output | 10-layer recurrent model · PTB | gradient flow across layers | Helped |
| Train with less data | LRU and competing recurrent models · PTB and text8 mini-datasets | test loss | Mixed |
| Increase training data from 80% to 100% | language models · text8 | test loss | Hurt |
Closest papers by meaning
- Resurrecting Recurrent Neural Networks for Long Sequences (ICML 2023 · similarity 0.88)
- Hierarchically Gated Recurrent Neural Network for Sequence Modeling (NeurIPS 2023 · similarity 0.88)
- LightRNN: Memory and Computation-Efficient Recurrent Neural Networks (NeurIPS 2016 · similarity 0.88)
- Stable Recurrent Models (ICLR 2019 · similarity 0.88)
- Recurrent Highway Networks (ICML 2017 · similarity 0.88)
- Combining Recurrent, Convolutional, and Continuous-time Models with Linear State Space Layers (NeurIPS 2021 · 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.