Reading about Machine Learning in Seismology
I’ve been slow to fully embrace machine learning in my research, but it’s rapidly becoming an important tool for seismological research. Recent papers in seismology, especially on the application of deep neural networks for automating seismic phase identification, picking, and association, have produced some really impressive results. In addition, it’s pretty clear that machine learning will have a role far beyond automation of classical data processing tasks, and into more advanced data-driven discovery, modeling, and inverse theory. So, admittedly somewhat reluctantly at first, I have begun to dive into the subfield of machine learning in seismology.
Reading the literature on machine learning in seismology is a challenge, however, because there’s a whole different terminology you’ve got to have at your fingertips. The architecture diagrams you see in these new types of seismic papers (which are mostly all just adapted from the computer science community) require at least some understanding of terms like ‘pooling’, ‘strides’, ‘skip connections’, ‘relu’, and many more. This vocabulary can’t be found in any seismology textbooks. So, to understand these papers, it seems like you suddenly have to be an expert in seismology *AND* machine learning, a tall order when keeping up with the advances in seismology is hard enough.
I’ve found two books to be particularly helpful to get me sufficiently up-to-speed on machine learning to tackle much of this literature. It’s been challenging to figure out how to approach this, because my goal is to follow and contribute to the seismology literature (not to simply run pre-built models at one extreme, or to become a machine learning specialist at the other). The first book – Deep Learning with Python by Francois Chollet – is a very practical one and explains things beautifully. I’ve found this book to be really great for getting a general understanding of the vocabulary and methods, but leaves me wanting a bit more understanding of the theory. For this, I’ve found Deep Learning by Goodfellow (the book most widely cited in the seismic literature) to be pretty good (but it works best for me in combination with Chollet). Goodfellow has more than I want in a lot of situations, but you can pick and choose which parts you want to know about and still get a lot out of it.
In my still relatively new job, I’m also thinking about education. The combination of disciplines required to get into this subfield is representative of what I think is one of the issues in academia today: the stove-piping of knowledge in specific disciplines. It really bugs me whenever I hear someone say: ‘to get a degree in XXXX, you really should have to take XXXX’. At a graduate level in particular, it seems like more-and-more, we should encourage diversity across the disciplines, as this is where many of the most important discoveries lie untapped. How can we help our students become more multidisciplinary, where our current academic model is designed to make students really good at one specific thing? It seems to me that this is an important problem we have to grapple with.