Thinking About Thinking: Five centuries of reductive science vs. artificial intelligence, with protein folding as an example

Tune in for the live stream on YouTube

Abstract: According to Aristotle “all, by nature, desire to know,” but how do we “know?” We are still learning how to learn about the physical universe. Since Galileo, the goal of scientific understanding is to explain complex phenomena with a compact description, a model, preferably one in which the description has physical meaning. For example, Tycho Brahe’s copious observations of planetary motions were reduced to Kepler’s three compact laws, an empirical mathematical description that was transformed into physics by Newton. This progression, from empirical data to abstract representation to a physical model, illustrates the ongoing, accretive process by which we learn. Newton knew this very well, writing famously that, “If I have seen further it is by standing on the shoulders of Giants” (Robert Hooke notwithstanding).

At the highest level, a model can predict surprising phenomena that have yet to be observed. Mendeleev’s periodic table of the elements, organized empirically by their characteristic properties, can be reproduced and completed using quantum chemistry, an entirely abstract theory. Could there be a more profound demonstration that scientific thinking is commensurate with the material world!

Yet today, artificial intelligence – specifically, machine-learning using neural nets – has engendered a radical departure from traditional approaches. In particular, reductionism, the cornerstone of scientific thinking since Galileo, is now poised at a transition point. The great scientific revolutions, relativity theory and quantum mechanics, ushered in profound paradigm shifts, but the conviction that an underlying theory could account for complex phenomena remained intact. Until now. Machine-learning using neural nets is not grounded in a unifying theory. There are no hypotheses being tested. Instead, the goal is to find parameters – often billions of parameters – that can capture the phenomenon under consideration and to then utilize the parameters predictively. It's pattern recognition: Keppler but not Newton. This approach has met with stunning success in multiple venues, but it is no longer science as we have come to know it. How shall we think about it? And where do we go from here? I will raise these questions, using the protein folding problem as an example.