Recently I was asked "is there benefit of reprocessing data and machine learning together?"

Yes. It has been standard business practice that every few years, with improved algorithm, we reprocess data, get higher resolution and a more detailed look. Like going from 4K to 8K HDTV, instead of 80 to 100 feet resolution in seismic, we may get that down to 40 ft. With higher resolution data, we’d retrain machine learning and get better results. Both go hand in hand.

That brings up a good point. In the world of geoscience, if we change the model, we also get different resulting imaged data. Unlike typical data used to feed machine learning algorithm, say what I bought from Amazon or movies streamed from Netflix, what I read and watched became record. That won’t change. But when imaging seismic, the model and resulting data are tightly coupled. Change one, we change the other.

In the machine learning world, we shall explore "reinforcement learning" in which we set guidelines what action to be taken and how we measure success. For precise definition, click on this link

https://en.wikipedia.org/...inforcement_learning

Perspective - think of what we do to map subsalt reservoirs. Depth imaging is an iterative and interpretive process. Update velocity-depth model with tomography, check for improved focusing and moveouts in the imaged gathers, then update salt geometry and velocity. If we do everything well, we converge to a better answer with each iteration. Reinforcement learning has the potential for productivity boost and automation of some of that work.

So here is food for thought on reinforcement learning.

Q1: What if by 2020, we can ask our digital assistant (be that Hey Google, Alexa, Cortana or Siri), for AI-first second opinion on the salt image we have just taken delivery?

Q2: how would that change our approach?

Feel free to chime in.