Paper Summary
An example of seismic data image processing of a real marine seismic dataset is used to demonstrate that Machine Learning has a future for seismic processing, but it also shows that huge training efforts are required for application to even modest input data sizes. As the applications of Machine Learning and Deep Neural Networks (DNNs) extend to ever-increasingly large and complex problems, the computational overheads will correspondingly become significant. Application-Specific Integrated Circuits will expectably become more common to address the vast computational challenges to training DNNs with modern 3D seismic data volumes, and will represent a new class of supercomputer with hundreds of PFLOPS of computing capacity available to solve each problem.