Complex Unitary Recurrent Neural Networks Using Scaled Cayley Transform
Authors: Kehelwala D. G. Maduranga, Kyle E. Helfrich, Qiang Ye4528-4535
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we compare the performances between the restricted-capacity u RNN, full-capacity u RNN, EURNN, LSTM, sco RNN, o RNN and scu RNN architectures on a variety of tasks. [...] Results of the experiments are given in Table 1, Figure 2, and Figure 3. |
| Researcher Affiliation | Academia | Kehelwala D. G. Maduranga, Kyle E. Helfrich, Qiang Ye Mathematics Department, University of Kentucky Lexington, KY, 40508, United States {kdgmaduranga,kyle.helfrich,qye3}@uky.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code for these experiments is available at https://github.com/Gayan225/scu RNN. |
| Open Datasets | Yes | MNIST database (Le Cun and Cortes 2010), TIMIT data set (Garofolo et al. 1993) |
| Dataset Splits | Yes | The core test set was used, consisting of a training set of 3,696 audio files, a testing set of 192 audio files, and a validation set of 400 audio files. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions optimizers (RMSProp, Adam, Adagrad) but does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | For each model, the hidden size was adjusted to match the number of trainable parameters. [...] The initial hidden state was initialized using the distribution U[−0.01, 0.01] and was trainable. [...] The θ values are sampled from U [0, 2π]. [...] The biases are initialized from the distribution U [−0.01, 0.01]. [...] We used several different combinations of optimizers and learning rates as noted under each experiment. [...] All the models were trained for a total of 70 epochs. |