13 April, 2022
A presentation titled "The geo/physics of and in neural networks" was given on 13th April 2022 in KAUST. The event was hosted as part of the KAUST Earth Science and Engineering Graduate Seminar series.
Neural networks, and the process behind training them, have been “labeled” by many physics scientists as anti-science, anti-physics, and even anti knowledge. Possibly rightfully so, since some data scientists have claimed that they will not need our scientific theories to solve physical problems any more. For decades, we have relied on our knowledge, experience, and developed tools to interpret data and predict the key geo/physical properties behind them. We, specifically, use numerical methods based on our physical laws to invert the data and obtain physical models of the Earth, as an example. So, machine learning, which is a numerical tool, but probably a more intrusive one, has been, fairly, accused of using the data to predict the physical properties, as simple “tasks”, with the physical knowledge rendered buried in the hidden layers (“the black box” phenomenon), nowhere to be found. My task in this presentation is to hopefully show you that training neural networks, reliant on a mix of statistics and inverse theory, can benefit heavily from our geo/physical laws and a priori knowledge for guidance, and specifically to alleviate the many weaknesses/gaps/biases in data. In fact, many of the neural network machinery have clear justifications in physical laws related to energy, entropy, and even Fermat’s principle. As an optimization framework in which physics and our prior knowledge can always be incorporated, and is actually needed, to solve complex physical problems, neural networks (NNs) definitely need geo/physical laws, and our physics-based challenges can benefit from the power of NNs. Examples in discovering the subsurface will shed light on this, probably well known, assertion of mine.