Recurrent Neural Network (RNN) has achieved many state-of-the-art performances on various complex tasks related to the temporal and sequential data. But most of these RNNs require much computational power and a huge number of parameters for both training and inference stage. Several tensor decomposition methods are included such as CANDECOMP/PARAFAC (CP), Tucker decomposition and Tensor Train (TT) to re-parameterize the Gated Recurrent Unit (GRU) RNN. First, we evaluate all tensor-based RNNs performance on sequence modeling tasks with a various number of parameters. Based on our experiment results, TT-GRU achieved the best results in a various number of parameters compared to other decomposition methods. Later, we evaluate our proposed TT-GRU with speech recognition task. We compressed the bidirectional GRU layers inside DeepSpeech2 architecture. Based on our experiment result, our proposed TT-format GRU are able to preserve the performance while reducing the number of GRU parameters significantly compared to the uncompressed GRU.
Andros TJANDRA
Nara Institute of Science and Technology,RIKEN, Center for Advanced Intelligence Project AIP
Sakriani SAKTI
Nara Institute of Science and Technology,RIKEN, Center for Advanced Intelligence Project AIP
Satoshi NAKAMURA
Nara Institute of Science and Technology,RIKEN, Center for Advanced Intelligence Project AIP
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Andros TJANDRA, Sakriani SAKTI, Satoshi NAKAMURA, "Recurrent Neural Network Compression Based on Low-Rank Tensor Representation" in IEICE TRANSACTIONS on Information,
vol. E103-D, no. 2, pp. 435-449, February 2020, doi: 10.1587/transinf.2019EDP7040.
Abstract: Recurrent Neural Network (RNN) has achieved many state-of-the-art performances on various complex tasks related to the temporal and sequential data. But most of these RNNs require much computational power and a huge number of parameters for both training and inference stage. Several tensor decomposition methods are included such as CANDECOMP/PARAFAC (CP), Tucker decomposition and Tensor Train (TT) to re-parameterize the Gated Recurrent Unit (GRU) RNN. First, we evaluate all tensor-based RNNs performance on sequence modeling tasks with a various number of parameters. Based on our experiment results, TT-GRU achieved the best results in a various number of parameters compared to other decomposition methods. Later, we evaluate our proposed TT-GRU with speech recognition task. We compressed the bidirectional GRU layers inside DeepSpeech2 architecture. Based on our experiment result, our proposed TT-format GRU are able to preserve the performance while reducing the number of GRU parameters significantly compared to the uncompressed GRU.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.2019EDP7040/_p
Copy
@ARTICLE{e103-d_2_435,
author={Andros TJANDRA, Sakriani SAKTI, Satoshi NAKAMURA, },
journal={IEICE TRANSACTIONS on Information},
title={Recurrent Neural Network Compression Based on Low-Rank Tensor Representation},
year={2020},
volume={E103-D},
number={2},
pages={435-449},
abstract={Recurrent Neural Network (RNN) has achieved many state-of-the-art performances on various complex tasks related to the temporal and sequential data. But most of these RNNs require much computational power and a huge number of parameters for both training and inference stage. Several tensor decomposition methods are included such as CANDECOMP/PARAFAC (CP), Tucker decomposition and Tensor Train (TT) to re-parameterize the Gated Recurrent Unit (GRU) RNN. First, we evaluate all tensor-based RNNs performance on sequence modeling tasks with a various number of parameters. Based on our experiment results, TT-GRU achieved the best results in a various number of parameters compared to other decomposition methods. Later, we evaluate our proposed TT-GRU with speech recognition task. We compressed the bidirectional GRU layers inside DeepSpeech2 architecture. Based on our experiment result, our proposed TT-format GRU are able to preserve the performance while reducing the number of GRU parameters significantly compared to the uncompressed GRU.},
keywords={},
doi={10.1587/transinf.2019EDP7040},
ISSN={1745-1361},
month={February},}
Copy
TY - JOUR
TI - Recurrent Neural Network Compression Based on Low-Rank Tensor Representation
T2 - IEICE TRANSACTIONS on Information
SP - 435
EP - 449
AU - Andros TJANDRA
AU - Sakriani SAKTI
AU - Satoshi NAKAMURA
PY - 2020
DO - 10.1587/transinf.2019EDP7040
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E103-D
IS - 2
JA - IEICE TRANSACTIONS on Information
Y1 - February 2020
AB - Recurrent Neural Network (RNN) has achieved many state-of-the-art performances on various complex tasks related to the temporal and sequential data. But most of these RNNs require much computational power and a huge number of parameters for both training and inference stage. Several tensor decomposition methods are included such as CANDECOMP/PARAFAC (CP), Tucker decomposition and Tensor Train (TT) to re-parameterize the Gated Recurrent Unit (GRU) RNN. First, we evaluate all tensor-based RNNs performance on sequence modeling tasks with a various number of parameters. Based on our experiment results, TT-GRU achieved the best results in a various number of parameters compared to other decomposition methods. Later, we evaluate our proposed TT-GRU with speech recognition task. We compressed the bidirectional GRU layers inside DeepSpeech2 architecture. Based on our experiment result, our proposed TT-format GRU are able to preserve the performance while reducing the number of GRU parameters significantly compared to the uncompressed GRU.
ER -