Cells in Multidimensional Recurrent Neural Networks
Authors: Gundram Leifert, Tobias Strauß, Tobias Grüning, Welf Wustlich, Roger Labahn
JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To test the performance of the cells in Section 8 we take two data sets from the ICDAR 2009 competitions, where the MDRNNs with LSTM cell won. On these data sets we compare the recognition results of the MDRNNs when we substitute the LSTM cells by the new developed cells. On both data sets, the IFN/ENIT data set and the RIMES data set we can improve the recognition rate using the new developed cells. |
| Researcher Affiliation | Academia | Gundram Leifert EMAIL Tobias Strauß EMAIL Tobias Grüning EMAIL Welf Wustlich EMAIL Roger Labahn EMAIL University of Rostock Institute of Mathematics 18051 Rostock, Germany |
| Pseudocode | No | The paper uses mathematical equations and schematic diagrams (e.g., Figure 1, 2, 3, 5) to describe the cell architectures and updates. However, it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code or provide links to a code repository. |
| Open Datasets | Yes | On both data sets, the IFN/ENIT data set and the RIMES data set we can improve the recognition rate using the new developed cells. (Pechwitz, M. and Maddouri, S. and Märgner, V. and Ellouze, N. and Amiri, H. and others, 2002) and the Rimes database (Augustin, E. and Brodin, J.-M and Carré, M. and Geoffrois, E. and Grosicki, E. and Prêteux, F., 2006). |
| Dataset Splits | Yes | Therefor we divide the sets a-e into 30000 training samples and 2493 validation samples. ... The Rimes database ... is divided into distinct subsets; a training set of 44196 samples, a validation set of 7542 samples and a test set of 7464 samples. |
| Hardware Specification | No | The paper mentions that RNNs were used for experiments but does not provide any specific hardware details such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions training networks with Backpropagation through time (BPTT) but does not specify any software libraries, frameworks (e.g., TensorFlow, PyTorch), or their version numbers. |
| Experiment Setup | Yes | All networks are trained 100 epochs with a fixed learning rate δ = 1 10 4. ... train all networks with stochastic gradient decent with different learning rates δ 1 10 3, 5 10 4, 2 10 4, 1 10 4 and compare the best LER according a fixed learning rate. ... The network used in this section differs only in the subsampling rate between two layers from the network used in A. Graves and J. Schmidhuber (2008). |