Seismic instruments: article

Deep artificial neural networks as a tool for the analysis of seismic data
K. Kislov
V. Gravirov
Institute of Earthquake Prediction Theory and Mathematical Geophysics, Russian Academy of Sciences, Moscow, Russia
Journal: Seismic instruments
Tome: 53
Number: 1
Year: 2017
Pages: 17-28
UDK: 004.32.26:550.34.64
DOI: 10.21455/si2017.1-2
Keywords: deep neural networks, deep learning, greedy algorithm, seismic data, multi-task learning
Аnnotation: The number of research in seismology based on an artificial neuron network technique has been increasing. However, the efficiency of neural networks with one hidden layer is limited. The last few years there has been a new neuroinformatics glut associated with the development of networks 3rd generation (deep neural networks). These networks operate with data at higher-level. The learning requires a smaller number of labeled data, and for pretraining of a network it is possible to use the unlabeled data. The networks have a higher level of abstraction and make fewer errors during the work. The deep network can be used for several tasks. The article is devoted to a possibility of application of deep networks in seismology. We are describing what the deep networks are; how to train them, what benefits they have, how to adapt them to the peculiarities of working with the seismic data, what prospects are opening up in connection with their application.