Research map / Architectures and efficiency / RNNs and normalization
The Statistical Recurrent Unit
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 SRU instead of gated recurrent units | Recurrent sequence models · Synthetic SRU-generated data | MSE: 0.62 vs 21.72 (GRU) and 161.62 (LSTM) | Helped |
| Use SRU instead of gated recurrent units | Recurrent sequence models · MNIST pixel sequence classification | error rate: 0.11 vs 0.28 (GRU) and 0.48 (LSTM) | Helped |
| Use SRU instead of GRU or LSTM | Polyphonic music modeling · JSB and Piano | negative log-likelihood: JSB: 8.260 vs 8.548 and 8.393; Piano: 7.737 vs 7.929 and 7.931 | Helped |
| Use SRU instead of GRU or LSTM | Polyphonic music modeling · Muse and Nottingham | negative log-likelihood: Muse: 6.336 vs LSTM 6.293; Nottingham: 3.362 vs LSTM 3.359 | Mixed |
| Use SRU instead of gated recurrent units | Recurrent sequence models · Electronica MFCC music | MSE: 1.176 vs 2.080 (GRU) and 1.183 (LSTM) | Helped |
| Use SRU instead of gated recurrent units | Recurrent sequence models · NARR climate data | MSE: 0.465 vs 0.487 (GRU) and 0.466 (LSTM) | Helped |
| Use SRU instead of gated recurrent units | Recurrent sequence models · SportVu NBA tracking | MSE: 34.505 vs 329.921 (GRU) and 296.908 (LSTM) | Helped |
| Add recurrent statistics and multiple scales rather than single-scale iid statistics (ablation) | SRU · MNIST pixel sequence classification | error rate: 0.88 (iid), 0.88 (recur), 0.63 (multi) | Helped |
| Add multiple scales rather than using only scales {0.0, 0.5, 0.9} (ablation) | SRU · MNIST pixel sequence classification | error rate: 0.79 -> 0.21 | Helped |
| Replace ReLU with tanh (ablation) | SRU · MNIST pixel sequence classification | error rate: 0.11 -> 0.18 | Hurt |
| Increase recurrent-summary dimensionality (ablation) | SRU · MNIST pixel sequence classification | error rate: 0.25 (r_dims 5), 0.20 (r_dims 20), 0.10 (r_dims 240) | Helped |
| Increase num_stats from 10 to 50 (ablation) | SRU · MNIST pixel sequence classification | error rate: 0.88 -> 0.32 | Helped |
Closest papers by meaning
- Towards Non-Saturating Recurrent Units for Modelling Long-Term Dependencies (AAAI 2019 · similarity 0.89)
- Preventing Gradient Explosions in Gated Recurrent Units (NeurIPS 2017 · similarity 0.88)
- Can recurrent neural networks warp time? (ICLR 2018 · similarity 0.87)
- Lattice Recurrent Unit: Improving Convergence and Statistical Efficiency for Sequence Modeling (AAAI 2018 · similarity 0.87)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.86)
- Real-Time Recurrent Learning using Trace Units in Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
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.