Explainable seismic neural networks using learning statistics

Abstract

PreviousNext No AccessFirst International Meeting for Applied Geoscience & Energy Expanded AbstractsExplainable seismic neural networks using learning statisticsAuthors: Ryan BenkertOluwaseun Joseph AribidoGhassan AlRegibRyan BenkertGeorgia Institute of TechnologySearch for more papers by this author, Oluwaseun Joseph AribidoGeorgia Institute of TechnologySearch for more papers by this author, and Ghassan AlRegibGeorgia Institute of TechnologySearch for more papers by this authorhttps://doi.org/10.1190/segam2021-3583675.1 SectionsAboutPDF/ePub ToolsAdd to favoritesDownload CitationsTrack CitationsPermissions ShareFacebookTwitterLinked InRedditEmail AbstractEven though deep neural networks are incredibly successful in speeding up seismic interpretation, they are frequently met with skepticism. Specifically, the main critique of deep learning remains the lack of explainable predictions and their poor generalization to difficult textures or structures that are not resent in the training volume. The objective of this paper is to address this highly relevant issue by explaining deep seismic networks with their behaviour during model training. Our approach is based on analysing how networks ”forget” previously learned seismic reflections and on visualizing the forgetting behavior in heat maps. We show that our method connects difficult seismic regions with incoherent neural network behaviour by establishing a link between forgettable seismic reflections and the learned decision boundary of neural networks. This makes our method especially useful for reliability and robustness analysis of deep seismic models. Finally, we group forgettable seismic regions into sub-categories by analyzing when the network forgets specific areas. Each subgroup is characterized by unique features that are related to the learning difficulty of specific subsurfaces and contain valuable insights in the interpretation properties of neural networks.Keywords: neural networks, machine learning, visualization, interpretationPermalink: https://doi.org/10.1190/segam2021-3583675.1FiguresReferencesRelatedDetails First International Meeting for Applied Geoscience & Energy Expanded AbstractsISSN (print):1052-3812 ISSN (online):1949-4645Copyright: 2021 Pages: 3561 publication data© 2021 Published in electronic format with permission by the Society of Exploration GeophysicistsPublisher:Society of Exploration Geophysicists HistoryPublished: 01 Sep 2021 CITATION INFORMATION Ryan Benkert, Oluwaseun Joseph Aribido, and Ghassan AlRegib, (2021), "Explainable seismic neural networks using learning statistics," SEG Technical Program Expanded Abstracts : 1425-1429. https://doi.org/10.1190/segam2021-3583675.1 Plain-Language Summary Keywordsneural networksmachine learningvisualizationinterpretationPDF DownloadLoading ...