Research map / Architectures and efficiency / RNNs and normalization
Improving the Performance and Learning Stability of Parallelizable RNNs Designed for Ultra-Low Power Applications
Julien Brandoit, Arthur Fyon, Damien Ernst, Guillaume Drion
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add cumulative update (epsilon 0 -> 1) (ablation) | CMRU d=32 · sMNIST | accuracy (%): 35.32 -> 95.92 | Improved |
| add cumulative update (epsilon 0 -> 1) (ablation) | alphaCMRU d=32 · sMNIST | accuracy (%): 48.86 -> 96.76 | Improved |
| anneal epsilon from 1 to 0 instead of constant epsilon=0 (ablation) | CMRU d=32 · sMNIST | accuracy (%): 35.32 -> 87.94 | Improved |
| train with constant epsilon=1 instead of annealed epsilon (ablation) | CMRU d=32 · sMNIST | accuracy (%): 87.94 -> 95.92 | Improved |
| anneal epsilon vs constant epsilon=1 (ablation) | alphaCMRU d=32 · sMNIST | accuracy (%): 96.76 -> 96.94 | No clear effect |
| use CMRU instead of LRU at d=4 | CMRU (single layer) · sMNIST | accuracy (%): 83.61 | Improved |
| use CMRU instead of LRU/minGRU for persistent memory | CMRU r=3 d=32 · Pathfinder | accuracy (%): 50.14 -> 90.20 | Improved |
| use CMRU instead of LRU on ListOps | CMRU r=3 d=32 · ListOps | accuracy (%): 40.45 -> 60.42 | Improved |
| use CMRU instead of LRU on IMDb | CMRU r=3 d=32 · IMDb | accuracy (%): 66.84 -> 65.18 | Hurt |
| use CMRU instead of minGRU on sCIFAR10 | CMRU r=3 d=32 · sCIFAR10 | accuracy (%): 60.95 -> 59.82 | Hurt |
| combine minGRU with CMRU layers | minGRU · Shakespeare char-level LM | cross-entropy loss: 1.453 -> 1.442 | Improved |
| use CMRU instead of LRU for keyword spotting | CMRU r=1 d=16 · KWS All | accuracy (%): 91.82 -> 89.76 | Hurt |
Closest papers by meaning
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.89)
- Learning Recurrent Binary/Ternary Weights (ICLR 2019 · similarity 0.88)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.88)
- Improving Bilinear RNN with Closed-loop Control (NeurIPS 2025 · similarity 0.87)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.87)
- Fixed-Point RNNs: Interpolating from Diagonal to Dense (NeurIPS 2025 · 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.
