Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Research › peer-review
Multi-output Deep Learning Framework for Solar Atmospheric Parameters Inferring from Stokes Profiles. / Knyazeva, Irina; Plotnikov, Andrey; Medvedeva, Tatiana; Makarenko, Nikolay.
Advances in Neural Computation, Machine Learning, and Cognitive Research 5 - Selected Papers from the 23rd International Conference on Neuroinformatics, 2021. ed. / Boris Kryzhanovsky; Witali Dunin-Barkowski; Vladimir Redko; Yury Tiumentsev; Valentin V. Klimov. Springer Nature, 2021. p. 299-307 (Studies in Computational Intelligence; Vol. 1008 SCI).Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Research › peer-review
}
TY - GEN
T1 - Multi-output Deep Learning Framework for Solar Atmospheric Parameters Inferring from Stokes Profiles
AU - Knyazeva, Irina
AU - Plotnikov, Andrey
AU - Medvedeva, Tatiana
AU - Makarenko, Nikolay
N1 - Publisher Copyright: © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - Spectropolarimetric observations are broadly used for the extraction of physical information in the field of solar physics. Inferring magnetic and thermodynamic information from these observations includes inversion problem solving. Assuming that spectropolarimetric profiles are produced by a given atmospheric model, it is required to find the best sets of parameters within such a model corresponding to particular observations. Standard optimization approach often requires large computational resources and even in this case still performs very slowly. Previously it was suggested to use different strategies with artificial neural networks to overcome problems with computational power. It was previously shown that neural networks could be a viable alternative to the standard least square approach, but they could not replace it. Most papers only cover Magnetic Fields Vector parameter inferring, whereas the commonly used solar atmosphere model includes 11 parameters. In this paper we provide an end-to-end deep learning framework for full parameter inferring as well as comparison of several approaches for multi-output predictions. For this purpose, we trained one common network to predict all parameters, a set of parameter-oriented independent networks to deal with each parameter, and finally a combination of the above: a set of parameter-oriented independent networks built upon several layers of the pretrained common network. Our results show that using a partly independent network built upon a pretrained network provides the best results and demonstrates better generalization performance.
AB - Spectropolarimetric observations are broadly used for the extraction of physical information in the field of solar physics. Inferring magnetic and thermodynamic information from these observations includes inversion problem solving. Assuming that spectropolarimetric profiles are produced by a given atmospheric model, it is required to find the best sets of parameters within such a model corresponding to particular observations. Standard optimization approach often requires large computational resources and even in this case still performs very slowly. Previously it was suggested to use different strategies with artificial neural networks to overcome problems with computational power. It was previously shown that neural networks could be a viable alternative to the standard least square approach, but they could not replace it. Most papers only cover Magnetic Fields Vector parameter inferring, whereas the commonly used solar atmosphere model includes 11 parameters. In this paper we provide an end-to-end deep learning framework for full parameter inferring as well as comparison of several approaches for multi-output predictions. For this purpose, we trained one common network to predict all parameters, a set of parameter-oriented independent networks to deal with each parameter, and finally a combination of the above: a set of parameter-oriented independent networks built upon several layers of the pretrained common network. Our results show that using a partly independent network built upon a pretrained network provides the best results and demonstrates better generalization performance.
KW - Deep learning
KW - Inverse problem
KW - Magnetic fields
KW - Multi-task learning
KW - Spectral lines
UR - http://www.scopus.com/inward/record.url?scp=85121582905&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-91581-0_40
DO - 10.1007/978-3-030-91581-0_40
M3 - Conference contribution
AN - SCOPUS:85121582905
SN - 9783030915803
T3 - Studies in Computational Intelligence
SP - 299
EP - 307
BT - Advances in Neural Computation, Machine Learning, and Cognitive Research 5 - Selected Papers from the 23rd International Conference on Neuroinformatics, 2021
A2 - Kryzhanovsky, Boris
A2 - Dunin-Barkowski, Witali
A2 - Redko, Vladimir
A2 - Tiumentsev, Yury
A2 - Klimov, Valentin V.
PB - Springer Nature
T2 - 23rd International Conference on Neuroinformatics, 2021
Y2 - 18 October 2021 through 22 October 2021
ER -
ID: 92451072