Computation is Learning
I worked in technology for years before I even gave the act of computation more than a second thought. Then about five years ago I went back to academia and started studying neuroscience, which led to quantitative chemistry and biology, and ultimately the beginning of what I think will be a decades-long research pursuit of how biology computes so efficiently and effectively. After a series of genetics and protein evolution, biochemistry, and neuro processing classes, I didn't really have a choice but to think more deeply about computation as a process. It seemed to me that silicon computation functioned in an extrapolated abstraction that was not at all how biology computed. But I am getting ahead of myself. Let's make sure we level set on the concepts we are discussing here so casually.
In my mind, computation has to have an objective. Adding two numbers is an abstraction unless you are adding two numbers for a purpose. Fundamentally, biology always has an objective, which is decreasing uncertainty through management of thermodynamic complexity. In other words, biology follows the laws of thermodynamics, which constantly pull everything towards its most stable, low-energy state. In practical terms, that means biology is using energy (thermodynamics), to decrease uncertainty, and that decrease in uncertainty is just another word for prediction. In neuroscience Karl Friston proposed a version of this as the free-energy principle. So biology uses its energy in order to predict its state. In computer programming terms it is truly a state machine. Wait, so why is prediction the point? Well, prediction is what enables survival. Let's not get more existential than that right now.
Outside biology, in countless tools leading up to computers, from reading tea leaves to watching the movement of the stars, humans have used ancillary tools to increase their ability to predict their environment. The development of external supporting computation was a magnificent leap in capability, but unlike biology, that computation is managed by a series of electrical signals forcing switches on and off, following instructions exactly as written, with memory and processing kept apart.
Now, let's leap to AI in its multitude of forms, and I am oversimplifying for the purposes of discussion, but most of the AI models we use are statistical models fit to the past. Meaning they must all be provided a series of data and information and in one way or another, they learn the importance of that information, called weighting, backwards in time. Said another way, by working backwards from their errors on historical data. This is known as backpropagation. Predictions are trained historically; they are not learned in real time. The difference in biology is that it takes what it has learned, stores it at different time scales, and constantly reweights that information as it encounters new information. In biology, computing and learning are the same act. Computation is learning.
Forced now to think about computation, I took a headlong deep dive into the math and physics being used in the study of how systems learn, how neural tissue computes, the energy-based models of machine learning, recursive estimation, and whole-Earth geophysical modeling and weather forecasting, which correct their models as new observations arrive. This led me to the disciplines of fields, graphs, topology, and predictive coding, the kind of error correction neuroscience proposes the brain uses to predict. What is remarkable about academia is that so many of the pieces we need to solve problems are out there, but the disciplines don't talk to each other, or don't understand their relevance to each other. In my next entry, I will delve into the components of our core engine and the types of problems we are using that physics and mathematics to solve.
Sarah Todd Davidson is the founder and CEO of PAUL Health and the founder of Eyrie. Subscribe on Substack.