BEGIN:VCALENDAR VERSION:2.0 METHOD:PUBLISH PRODID:-//Tendenci - The Open Source AMS for Associations//Tendenci Codeba se MIMEDIR//EN BEGIN:VEVENT DESCRIPTION:--- This iCal file does *NOT* confirm registration.\r\nEvent d etails subject to change. ---\r\nhttps://www.spegcs.org/events/4206/\r\n\r \nEvent Title: Data Analytics Seminar: Seismic Applications of Machine Lea rning\r\nStart Date / Time: Jan 23, 2019 18:00 PM US/Central\r\nLocation: Intelie - Rignet Office\r\nSpeaker: Stephen Alwon, Sribharath Kainkaryam\r \nGoogle\r\nhttp://maps.google.com/maps?q=15115+Park+Row+Dr.+Ste+300+,Hous ton,Texas,77084\r\n\r\nForecast\nhttp://www.weather.com/weather/monthly/77 084\r\n\r\n \r\n \r\nGenerative adversarial networks in seismic data proce ssing (Stephen Alwon, Schlumberger)\r\nGenerative adversarial networks (GA Ns) are a class of machine learning techniques that involve two networks t rained simultaneously to generate a desired outcome. These schemes have ha d success in many traditional image processing tasks, such as style transf er and super-resolution, but are relatively unexplored in geophysics. We o utline the underlying theory behind GANs and present networks that can per form traditional seismic processing tasks such as noise attenuation and tr ace interpolation.\r\nConvolutional neural network for Salt model building in the Gulf of Mexico (Sribharath Kainkaryam, TGS )\r\nSalt model buildin g is one of the most time consuming, labor intensive and difficult to auto mate processes in the entire depth imaging workflow requiring significant human effort. The challenge and need for automating salt interpretation is well recognized by the seismic imaging community and has seen many applic ations of deep learning based convolutional neural network (CNN) architect ures to carry out this task. However, significant challenges remain for re liable production-scale deployment of CNN based methods for salt model bui lding owing to poor generalization capabilities of these networks. When us ed on new surveys, never seen by the CNN models during the training stage, the interpretation accuracy of these models drops significantly. To remed iate this key problem, we introduce a U-shaped encoder-decoder type CNN ar chitecture and use a specialized regularization strategy aimed at reducing the generalization error of the network. Our regularization scheme pertur bs the ground truth labels in the training set. Two different perturbation s are discussed: one which randomly changes the labels of the training set , flipping salt labels to sediments and vice versa and the second which sm ooths the labels. We demonstrate that such perturbations act as a strong regularizer preventing the network from making low entropy, highly confide nt predictions on the training set and thus reducing overfitting. An ensem ble strategy is also devised for test time augmentation that is shown to f urther improve the accuracy. The robustness, in terms of vastly improved g eneralization capability as well as improved interpretation accuracy of ou r training strategy is demonstrated with real data examples from the Gulf of Mexico. \r\n \r\nSponsors:\r\n \r\n --- This iCal file does *NOT* confi rm registration.Event details subject to change. ---\r\n\r\n--- By Tendenc i - The Open Source AMS for Associations ---\r\n UID:uid4206@spegcs.org SUMMARY:Data Analytics Seminar: Seismic Applications of Machine Learning DTSTART:20190124T000000Z DTEND:20190124T013000Z CLASS:PUBLIC PRIORITY:5 DTSTAMP:20240328T202117Z TRANSP:OPAQUE SEQUENCE:0 LOCATION:Intelie - Rignet Office X-ALT-DESC;FMTTYPE=text/html:
Generative adversarial networks (G ANs) are a class of machine learning techniques that involve two networks trained simultaneously to generate a desired outcome. These schemes have h ad success in many traditional image processing tasks, such as style trans fer and super-resolution, but are relatively unexplored in geophysics. We outline the underlying theory behind GANs and present networks that can pe rform traditional seismic processing tasks such as noise attenuation and t race interpolation.
Salt model bui lding is one of the most time consuming, labor intensive and difficult to automate processes in the entire depth imaging workflow requiring signific ant human effort. The challenge and need for automating salt interpretatio n is well recognized by the seismic imaging community and has seen many ap plications of deep learning based convolutional neural network (CNN) archi tectures to carry out this task. However, significant challenges remain fo r reliable production-scale deployment of CNN based methods for salt model building owing to poor generalization capabilities of these networks. Whe n used on new surveys, never seen by the CNN models during the training st age, the interpretation accuracy of these models drops significantly. To r emediate this key problem, we introduce a U-shaped encoder-decoder type CN N architecture and use a specialized regularization strategy aimed at redu cing the generalization error of the network. Our regularization scheme pe rturbs the ground truth labels in the training set. Two different perturba tions are discussed: one which randomly changes the labels of the training set, flipping salt labels to sediments and vice versa and the second whic h smooths the labels. \; We demonstrate that such perturbations act as a strong regularizer preventing the network from making low entropy, high ly confident predictions on the training set and thus reducing overfitting . An ensemble strategy is also devised for test time augmentation that is shown to further improve the accuracy. The robustness, in terms of vastly improved generalization capability as well as improved interpretation accu racy of our training strategy is demonstrated with real data examples from the Gulf of Mexico. \;
 \;
Sponsors:
 \;
 \;