Understanding Convolutional Neural Networks for Seismic Features by Visualization

In recent years, Convolutional Neural Networks (CNNs), that are most commonly applied to visual image analysis, have yielded astonishing results in a number of classification tasks. The areas of application include the recognition of text, speech and various objects and structures. However, the various activities inside of CNNs and their reactions to specific inputs remain unclear. For a better understanding of processes within a CNN, correlations between neurons and layers have to be visualized to be able to compare the reaction to different inputs afterwards. There are various approaches to visualize activities in CNNs in order to strengthen a users confidence in the classification results by offering a look inside the networks that usually work like black-boxes. In this thesis, five visualization methods (Activation Maximization, Deconvolution, Guided Backpropagation, Grad-CAM and Guided Grad-CAM) are applied to CNN architectures that are trained to detect seismic features in geological data. By visualizing features and other information from these CNNs, it is possible to gain knowledge about the decision-making process in the network and even derive optimization possibilities.