Research map / Architectures and efficiency / RNNs and normalization
Towards Non-Saturating Recurrent Units for Modelling Long-Term Dependencies
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 NRU instead of GRU | Character-level language modelling · PTB test | BPC: 1.45 -> 1.47 | Hurt |
| Use NRU instead of GRU | Character-level language modelling · PTB test | Accuracy: 69.07 -> 68.48 | Hurt |
| Use NRU instead of EURNN | Recurrent model · psMNIST | Accuracy: 94.50 -> 95.38 | Helped |
| Replace additive memory updates with multiplicative updates (ablation) | NRU · PTB character-level language modelling | BPC | Helped |
| Use ReLU writing and erasing heads instead of linear heads (ablation) | NRU · Most experiments other than language modelling | Performance | Hurt |
| Use ReLU writing and erasing heads instead of linear heads (ablation) | NRU · Character-level language modelling | Performance | Helped |
| Increase the number of memory heads (ablation) | NRU · psMNIST validation | Accuracy | Helped |
| Increase memory size from 256 to 400 (ablation) | NRU · psMNIST validation | Accuracy | No clear effect |
| Reduce hidden-state size from 200 to 50 or 100 (ablation) | NRU · psMNIST validation | Accuracy | Hurt |
| Increase memory size from 64 to 144 (ablation) | NRU · Copy task | Convergence speed: almost twice as early | Helped |
| Use chrono-initialization with T_max 50 instead of 150 (ablation) | LSTM and JANET · PTB character-level language modelling | BPC | Helped |
Closest papers by meaning
- Can recurrent neural networks warp time? (ICLR 2018 · similarity 0.89)
- Learning Long Term Dependencies via Fourier Recurrent Units (ICML 2018 · similarity 0.89)
- The Statistical Recurrent Unit (ICML 2017 · similarity 0.89)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.88)
- Efficient recurrent architectures through activity sparsity and sparse back-propagation through time (ICLR 2023 · similarity 0.88)
- Training biologically plausible recurrent neural networks on cognitive tasks with long-term dependencies (NeurIPS 2023 · 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.