On the initialization of long short-term memory networks

Publikation: Bidrag til bog/antologi/rapportKonferencebidrag i proceedingsForskningfagfællebedømt

Standard

On the initialization of long short-term memory networks. / Mehdipour Ghazi, Mostafa; Nielsen, Mads; Pai, Akshay; Modat, Marc; Cardoso, M. Jorge; Ourselin, Sébastien; Sørensen, Lauge.

Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings. red. / Tom Gedeon; Kok Wai Wong; Minho Lee. Springer VS, 2019. s. 275-286 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Bind 11953 LNCS).

Publikation: Bidrag til bog/antologi/rapportKonferencebidrag i proceedingsForskningfagfællebedømt

Harvard

Mehdipour Ghazi, M, Nielsen, M, Pai, A, Modat, M, Cardoso, MJ, Ourselin, S & Sørensen, L 2019, On the initialization of long short-term memory networks. i T Gedeon, KW Wong & M Lee (red), Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings. Springer VS, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), bind 11953 LNCS, s. 275-286, 26th International Conference on Neural Information Processing, ICONIP 2019, Sydney, Australien, 12/12/2019. https://doi.org/10.1007/978-3-030-36708-4_23

APA

Mehdipour Ghazi, M., Nielsen, M., Pai, A., Modat, M., Cardoso, M. J., Ourselin, S., & Sørensen, L. (2019). On the initialization of long short-term memory networks. I T. Gedeon, K. W. Wong, & M. Lee (red.), Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings (s. 275-286). Springer VS. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Bind. 11953 LNCS https://doi.org/10.1007/978-3-030-36708-4_23

Vancouver

Mehdipour Ghazi M, Nielsen M, Pai A, Modat M, Cardoso MJ, Ourselin S o.a. On the initialization of long short-term memory networks. I Gedeon T, Wong KW, Lee M, red., Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings. Springer VS. 2019. s. 275-286. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Bind 11953 LNCS). https://doi.org/10.1007/978-3-030-36708-4_23

Author

Mehdipour Ghazi, Mostafa ; Nielsen, Mads ; Pai, Akshay ; Modat, Marc ; Cardoso, M. Jorge ; Ourselin, Sébastien ; Sørensen, Lauge. / On the initialization of long short-term memory networks. Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings. red. / Tom Gedeon ; Kok Wai Wong ; Minho Lee. Springer VS, 2019. s. 275-286 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Bind 11953 LNCS).

Bibtex

@inproceedings{9361a5045c914564ab1548a8d4158aaf,
title = "On the initialization of long short-term memory networks",
abstract = "Weight initialization is important for faster convergence and stability of deep neural networks training. In this paper, a robust initialization method is developed to address the training instability in long short-term memory (LSTM) networks. It is based on a normalized random initialization of the network weights that aims at preserving the variance of the network input and output in the same range. The method is applied to standard LSTMs for univariate time series regression and to LSTMs robust to missing values for multivariate disease progression modeling. The results show that in all cases, the proposed initialization method outperforms the state-of-the-art initialization techniques in terms of training convergence and generalization performance of the obtained solution.",
keywords = "Deep neural networks, Disease progression modeling, Initialization, Long short-term memory, Time series regression",
author = "{Mehdipour Ghazi}, Mostafa and Mads Nielsen and Akshay Pai and Marc Modat and Cardoso, {M. Jorge} and S{\'e}bastien Ourselin and Lauge S{\o}rensen",
year = "2019",
doi = "10.1007/978-3-030-36708-4_23",
language = "English",
isbn = "9783030367077",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer VS",
pages = "275--286",
editor = "Tom Gedeon and Wong, {Kok Wai} and Minho Lee",
booktitle = "Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings",
note = "26th International Conference on Neural Information Processing, ICONIP 2019 ; Conference date: 12-12-2019 Through 15-12-2019",

}

RIS

TY - GEN

T1 - On the initialization of long short-term memory networks

AU - Mehdipour Ghazi, Mostafa

AU - Nielsen, Mads

AU - Pai, Akshay

AU - Modat, Marc

AU - Cardoso, M. Jorge

AU - Ourselin, Sébastien

AU - Sørensen, Lauge

PY - 2019

Y1 - 2019

N2 - Weight initialization is important for faster convergence and stability of deep neural networks training. In this paper, a robust initialization method is developed to address the training instability in long short-term memory (LSTM) networks. It is based on a normalized random initialization of the network weights that aims at preserving the variance of the network input and output in the same range. The method is applied to standard LSTMs for univariate time series regression and to LSTMs robust to missing values for multivariate disease progression modeling. The results show that in all cases, the proposed initialization method outperforms the state-of-the-art initialization techniques in terms of training convergence and generalization performance of the obtained solution.

AB - Weight initialization is important for faster convergence and stability of deep neural networks training. In this paper, a robust initialization method is developed to address the training instability in long short-term memory (LSTM) networks. It is based on a normalized random initialization of the network weights that aims at preserving the variance of the network input and output in the same range. The method is applied to standard LSTMs for univariate time series regression and to LSTMs robust to missing values for multivariate disease progression modeling. The results show that in all cases, the proposed initialization method outperforms the state-of-the-art initialization techniques in terms of training convergence and generalization performance of the obtained solution.

KW - Deep neural networks

KW - Disease progression modeling

KW - Initialization

KW - Long short-term memory

KW - Time series regression

UR - http://www.scopus.com/inward/record.url?scp=85077503119&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-36708-4_23

DO - 10.1007/978-3-030-36708-4_23

M3 - Article in proceedings

AN - SCOPUS:85077503119

SN - 9783030367077

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 275

EP - 286

BT - Neural Information Processing - 26th International Conference, ICONIP 2019, Proceedings

A2 - Gedeon, Tom

A2 - Wong, Kok Wai

A2 - Lee, Minho

PB - Springer VS

T2 - 26th International Conference on Neural Information Processing, ICONIP 2019

Y2 - 12 December 2019 through 15 December 2019

ER -

ID: 237712952