Research map / Architectures and efficiency / RNNs and normalization
Complex Unitary Recurrent Neural Networks Using Scaled Cayley Transform
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 a complex scaled Cayley transform with trainable diagonal phases instead of a fixed discrete scaling matrix | scuRNN · MNIST | test accuracy: 0.973 -> 0.976 | Helped |
| Use a complex scaled Cayley transform with trainable diagonal phases instead of a fixed discrete scaling matrix | scuRNN · Permuted MNIST | test accuracy: 0.943 -> 0.949 | Helped |
| Use a complex scaled Cayley recurrent matrix | scuRNN · TIMIT validation | MSE: 4.76 -> 3.94; 2.64 -> 1.84 | Helped |
| Use a complex scaled Cayley recurrent matrix | scuRNN · TIMIT evaluation | MSE: 4.31 -> 3.56; 2.39 -> 1.67 | Helped |
| Use scuRNN rather than restricted-capacity uRNN or LSTM | Copying problem · T=2000 | cross entropy | Helped |
| Use scuRNN rather than scoRNN | Copying problem · T=1000 | cross entropy | No clear effect |
| Use scuRNN rather than scoRNN | Adding problem · T=200 | MSE | Helped |
| Use scuRNN rather than scoRNN | Adding problem · T=750 | MSE | Mixed |
| Constrain modReLU bias to remain nonpositive (ablation) | Complex unitary RNNs · MNIST | test accuracy | Hurt |
| Use RMSProp for A and D and other weights instead of the mixed optimizer configuration (ablation) | scuRNN · MNIST | test accuracy: 0.976 -> 0.956 | Hurt |
| Increase the learning rate for A from 10^-4 to 10^-3, using RMSProp for A, D, and other weights (ablation) | scuRNN · MNIST | test accuracy: 0.956 -> 0.913 | Hurt |
| Set D's RMSProp learning rate to 10^-4 rather than 10^-3, with A at 10^-4 and other weights at 10^-3 (ablation) | scuRNN · MNIST | test accuracy: 0.956 -> 0.976 | No clear effect |
Closest papers by meaning
- Orthogonal Recurrent Neural Networks with Scaled Cayley Transform (ICML 2018 · similarity 0.96)
- Unitary Evolution Recurrent Neural Networks (ICML 2016 · similarity 0.89)
- Full-Capacity Unitary Recurrent Neural Networks (NeurIPS 2016 · similarity 0.89)
- Improved memory in recurrent neural networks with sequential non-normal dynamics (ICLR 2020 · similarity 0.88)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.86)
- Learning Useful Representations of Recurrent Neural Network Weight Matrices (ICML 2024 · similarity 0.85)
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.