Standard
Direct training of dynamic observation noise with UMarineNet. / Oehmcke, Stefan; Zielinski, Oliver; Kramer, Oliver.
Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings. ed. / Vera Kurkova; Barbara Hammer; Yannis Manolopoulos; Lazaros Iliadis; Ilias Maglogiannis. Springer Verlag, 2018. p. 123-133 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 11139 LNCS).
Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Harvard
Oehmcke, S, Zielinski, O & Kramer, O 2018,
Direct training of dynamic observation noise with UMarineNet. in V Kurkova, B Hammer, Y Manolopoulos, L Iliadis & I Maglogiannis (eds),
Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings. Springer Verlag, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11139 LNCS, pp. 123-133, 27th International Conference on Artificial Neural Networks, ICANN 2018, Rhodes, Greece,
04/10/2018.
https://doi.org/10.1007/978-3-030-01418-6_13
APA
Oehmcke, S., Zielinski, O., & Kramer, O. (2018).
Direct training of dynamic observation noise with UMarineNet. In V. Kurkova, B. Hammer, Y. Manolopoulos, L. Iliadis, & I. Maglogiannis (Eds.),
Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings (pp. 123-133). Springer Verlag,. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) Vol. 11139 LNCS
https://doi.org/10.1007/978-3-030-01418-6_13
Vancouver
Oehmcke S, Zielinski O, Kramer O.
Direct training of dynamic observation noise with UMarineNet. In Kurkova V, Hammer B, Manolopoulos Y, Iliadis L, Maglogiannis I, editors, Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings. Springer Verlag,. 2018. p. 123-133. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 11139 LNCS).
https://doi.org/10.1007/978-3-030-01418-6_13
Author
Oehmcke, Stefan ; Zielinski, Oliver ; Kramer, Oliver. / Direct training of dynamic observation noise with UMarineNet. Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings. editor / Vera Kurkova ; Barbara Hammer ; Yannis Manolopoulos ; Lazaros Iliadis ; Ilias Maglogiannis. Springer Verlag, 2018. pp. 123-133 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 11139 LNCS).
Bibtex
@inproceedings{c9b015e8f04440528a63c2a641f6bf43,
title = "Direct training of dynamic observation noise with UMarineNet",
abstract = "Accurate uncertainty predictions are crucial to assess the reliability of a model, especially for neural networks. Part of this uncertainty is the observation noise, which is dynamic in our marine virtual sensor task. Typically, dynamic noise is not trained directly, but approximated through terms in the loss function. Unfortunately, this noise loss function needs to be scaled by a trade-off-parameter to achieve accurate uncertainties. In this paper we propose an upgrade to the existing architecture, which increases interpretability and introduces a novel direct training procedure for dynamic noise modelling. To that end, we train the point prediction model and the noise model separately. We present a new loss function that requires Monte Carlo runs of the model to directly train for the uncertainty prediction accuracy. In an experimental evaluation, we show that in most tested cases the uncertainty prediction is more accurate than the manually tuned trade-off-parameter. Because of the architectural changes we are able to analyze the importance of individual parts of the time series of our prediction.",
keywords = "CNN, LSTM, Predictive uncertainty, Time series",
author = "Stefan Oehmcke and Oliver Zielinski and Oliver Kramer",
year = "2018",
month = jan,
day = "1",
doi = "10.1007/978-3-030-01418-6_13",
language = "English",
isbn = "9783030014179",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer Verlag,",
pages = "123--133",
editor = "Vera Kurkova and Barbara Hammer and Yannis Manolopoulos and Lazaros Iliadis and Ilias Maglogiannis",
booktitle = "Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings",
note = "27th International Conference on Artificial Neural Networks, ICANN 2018 ; Conference date: 04-10-2018 Through 07-10-2018",
}
RIS
TY - GEN
T1 - Direct training of dynamic observation noise with UMarineNet
AU - Oehmcke, Stefan
AU - Zielinski, Oliver
AU - Kramer, Oliver
PY - 2018/1/1
Y1 - 2018/1/1
N2 - Accurate uncertainty predictions are crucial to assess the reliability of a model, especially for neural networks. Part of this uncertainty is the observation noise, which is dynamic in our marine virtual sensor task. Typically, dynamic noise is not trained directly, but approximated through terms in the loss function. Unfortunately, this noise loss function needs to be scaled by a trade-off-parameter to achieve accurate uncertainties. In this paper we propose an upgrade to the existing architecture, which increases interpretability and introduces a novel direct training procedure for dynamic noise modelling. To that end, we train the point prediction model and the noise model separately. We present a new loss function that requires Monte Carlo runs of the model to directly train for the uncertainty prediction accuracy. In an experimental evaluation, we show that in most tested cases the uncertainty prediction is more accurate than the manually tuned trade-off-parameter. Because of the architectural changes we are able to analyze the importance of individual parts of the time series of our prediction.
AB - Accurate uncertainty predictions are crucial to assess the reliability of a model, especially for neural networks. Part of this uncertainty is the observation noise, which is dynamic in our marine virtual sensor task. Typically, dynamic noise is not trained directly, but approximated through terms in the loss function. Unfortunately, this noise loss function needs to be scaled by a trade-off-parameter to achieve accurate uncertainties. In this paper we propose an upgrade to the existing architecture, which increases interpretability and introduces a novel direct training procedure for dynamic noise modelling. To that end, we train the point prediction model and the noise model separately. We present a new loss function that requires Monte Carlo runs of the model to directly train for the uncertainty prediction accuracy. In an experimental evaluation, we show that in most tested cases the uncertainty prediction is more accurate than the manually tuned trade-off-parameter. Because of the architectural changes we are able to analyze the importance of individual parts of the time series of our prediction.
KW - CNN
KW - LSTM
KW - Predictive uncertainty
KW - Time series
UR - http://www.scopus.com/inward/record.url?scp=85054878493&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-01418-6_13
DO - 10.1007/978-3-030-01418-6_13
M3 - Article in proceedings
AN - SCOPUS:85054878493
SN - 9783030014179
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 123
EP - 133
BT - Artificial Neural Networks and Machine Learning – ICANN 2018 - 27th International Conference on Artificial Neural Networks, 2018, Proceedings
A2 - Kurkova, Vera
A2 - Hammer, Barbara
A2 - Manolopoulos, Yannis
A2 - Iliadis, Lazaros
A2 - Maglogiannis, Ilias
PB - Springer Verlag,
T2 - 27th International Conference on Artificial Neural Networks, ICANN 2018
Y2 - 4 October 2018 through 7 October 2018
ER -
