State-Frequency Memory Recurrent Neural Networks
Authors: Hao Hu, Guo-Jun Qi
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations on several temporal modeling tasks demonstrate the SFM can yield competitive performances, in particular as compared with the state-of-the-art LSTM models. |
| Researcher Affiliation | Academia | 1University of Central Florida, Orlando, FL, USA. Correspondence to: Guo-Jun Qi <Guojun.Qi@ucf.edu>. |
| Pseudocode | No | No pseudocode or algorithm blocks were found. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code or a link to a repository for the described methodology. |
| Open Datasets | Yes | The experimental results are obtained on four polyphonic music benchmarks that have been used in (Boulanger-lewandowski et al., 2012): Muse Data, JSB chorales (Allan & Williams, 2004), Piano-midi.de (Poliner & Ellis, 2007) and Nottingham. We perform the classification task on TIMIT speech corpus (Garofolo et al., 1993). |
| Dataset Splits | Yes | In experiments, we randomly select 800 sequences per type for training and the remaining are for testing. We follow the same protocol (Boulanger-lewandowski et al., 2012) to split the training and test set to make a fair comparison on the four datasets. In addition, we randomly select 184 utterances from the training set as the validation set and keep the rest for training. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments. |
| Software Dependencies | No | We implement the proposed SFM model using Theano Python Math Library (Team et al., 2016). (No specific version numbers for Python or Theano are provided.) |
| Experiment Setup | Yes | Unless otherwise specified, we train all the networks through the BPTT algorithm with the Ada Delta optimizer (Zeiler, 2012), where the decay rate is set to 0.95. All the weights are randomly initialized in the range [ 0.1, 0.1] and the learning rate is set to 10 4. The training objective is to minimize the frame level cross-entropy loss. |