Research map / Architectures and efficiency / RNNs and normalization
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
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 EURNN tunable capacity 2 instead of LSTM with the same parameter count | Pixel-permuted MNIST · Pixel-permuted MNIST test | accuracy: 0.902 -> 0.937 | Helped |
| Use EURNN tunable capacity 2 instead of URNN | Pixel-permuted MNIST · Pixel-permuted MNIST test | accuracy: 0.933 -> 0.937 | Helped |
| Use FFT-style unitary factorization instead of tunable capacity 2 (ablation) | EURNN · Pixel-permuted MNIST test | accuracy: 0.937 -> 0.925 | Hurt |
| Use EURNN tunable capacity 2 instead of LSTM with hidden size 64 | TIMIT · TIMIT test | MSE: 66.0 -> 63.3 | Helped |
| Use EURNN tunable capacity 2 instead of LSTM with hidden size 128 | TIMIT · TIMIT test | MSE: 54.5 -> 63.3 | Hurt |
| Increase unitary capacity from 2 to 32 (ablation) | EURNN · TIMIT test | MSE: 63.3 -> 52.7 | Helped |
| Increase unitary capacity from 2 to 128 (ablation) | EURNN · TIMIT test | MSE: 63.3 -> 51.9 | Helped |
| Use FFT-style unitary factorization instead of tunable capacity 2 (ablation) | EURNN · TIMIT test | MSE: 63.3 -> 52.4 | Helped |
| Use full-capacity EURNN rather than a small-capacity EURNN (ablation) | EURNN · TIMIT speech prediction | MSE: 63.3 -> 51.9 | Helped |
| Use a smaller-capacity or FFT-style EURNN rather than a full-capacity unitary matrix (ablation) | EURNN · Copying task, T=1000 | cross entropy | Helped |
Closest papers by meaning
- Full-Capacity Unitary Recurrent Neural Networks (NeurIPS 2016 · similarity 0.90)
- Unitary Evolution Recurrent Neural Networks (ICML 2016 · similarity 0.88)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.87)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.85)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.85)
- Learning Useful Representations of Recurrent Neural Network Weight Matrices (ICML 2024 · similarity 0.84)
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.