Research map / Architectures and efficiency / RNNs and normalization
Full-Capacity Unitary Recurrent Neural Networks
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 |
|---|---|---|---|
| Optimize the recurrence over all unitary matrices rather than using the restricted parameterization | uRNN · System identification with N>7 | best normalized MSE: 3.51\text{e}{-1} -> 5.04\text{e}{-2} (W_u, N=8); 7.60\text{e}{-1} -> 1.00\text{e}{-6} (W_g, N=16) | Helped |
| Optimize the recurrence over all unitary matrices rather than using the restricted parameterization | uRNN · System identification with N<7 | best normalized MSE: 4.81\text{e}{-1} -> 1.28\text{e}{-1} (W_u, N=4); 3.21\text{e}{-4} -> 8.72\text{e}{-2} (W_g, N=4) | Mixed |
| Use full-capacity recurrence instead of restricted-capacity recurrence | uRNN · Copy memory T=1000 and T=2000 | average cross entropy: 0.020 -> zero (T=1000); 0.010 -> zero (T=2000) | Helped |
| Use full-capacity recurrence instead of restricted-capacity recurrence | uRNN · TIMIT evaluation, approximately 83k parameters | Eval. MSE: 14.87 -> 15.24 | Mixed |
| Use full-capacity recurrence instead of restricted-capacity recurrence | uRNN · TIMIT evaluation, approximately 135k parameters | Eval. MSE: 15.17 -> 14.66 | Helped |
| Use full-capacity recurrence instead of restricted-capacity recurrence | uRNN · TIMIT evaluation, approximately 200k parameters | Eval. MSE: 16.14 -> 14.45 | Helped |
| Use a full-capacity uRNN instead of an LSTM | speech frame prediction · TIMIT evaluation, approximately 200k parameters | Eval. MSE: 15.80 -> 14.45 | Helped |
| Use a full-capacity uRNN instead of an LSTM | pixel-by-pixel MNIST · Permuted evaluation | accuracy: 91.7 -> 94.1 | Helped |
| Use a full-capacity uRNN instead of an LSTM | pixel-by-pixel MNIST · Unpermuted evaluation | accuracy: 98.2 -> 96.9 | Hurt |
| Increase full-capacity uRNN hidden size from N=116 to N=512 | pixel-by-pixel MNIST · Unpermuted evaluation | accuracy: 92.8 -> 96.9 | Helped |
| Increase full-capacity uRNN hidden size from N=116 to N=512 | pixel-by-pixel MNIST · Permuted evaluation | accuracy: 92.1 -> 94.1 | Helped |
Closest papers by meaning
- Unitary Evolution Recurrent Neural Networks (ICML 2016 · similarity 0.91)
- Learning Unitary Operators with Help From u(n) (AAAI 2017 · similarity 0.90)
- Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs (ICML 2017 · similarity 0.90)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.88)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.87)
- Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvalues (ICML 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.