Wisdom without a brain

Thinking through my preparation of a lecture about homeostasis

With the semester winding down and my latest dive into class live blogging coming to an end, I’m plotting out the next course for the blog. Looking at my Notes App, I’m guessing it may be the “potpourri” column for the next while. To choose from the rather scattered assortment of directions, I’m going with reactive mode, as I am a bit panicked about putting together a talk for tomorrow. Hopefully, rambling on the blog will get my thoughts in order.

This week, I’m attending a workshop at the Simons Institute on “Modern Paradigms in Generalization.” Narrowly construed, generalization mathematically characterizes how well a machine learning model performs on the next data point based on properties of data in the past. This typically means you assume the future is the same as the past and apply some contorted manipulation of the law of large numbers to get a prediction interval around your machine learning model predictions.

This workshop tries to get after something bigger. For the organizers, “Generalization, broadly construed, is the ability of machine learning methods to perform well in scenarios outside their training data.” That’s pretty broadly construed! They further propose:

We will characterize not just what it means for a machine learning model to do well on future data, but more generally for any entity to behave effectively in unknown future settings.

Ambitious!

The organizers kindly invited me to speak but told me I could talk about whatever I wanted. After consulting some of the other Simons program participants, I decided to go after the second half of that sentence. I’ve been grappling with a concept for the last year, a concept that precisely captures this characterization while looking nothing like artificial intelligence: homeostasis.

Homeostasis is the term coined by Walter Cannon to describe how organisms maintain some notion of internal constancy despite existing in an extremely adversarial environment. In his 1932 text, The Wisdom of the Body, Cannon deliberately chose homeostasis over equilibrium:

“The constant conditions which are maintained in the body might be termed equilibria. That word, however, has come to have fairly exact meaning as applied to relatively simple physico-chemical states, in closed systems, where known forces are balanced. The coordinated physiological processes which maintain most of the steady states in the organism are so complex and so peculiar to living beings–involving, as they may, the brain and nerves, the heart, lungs, kidneys, and spleen, all working cooperatively—that I have suggested a special designation for these states, homeostasis. The word does not imply something set and immobile, a stagnation. It means a condition—a condition which may vary, but which is relatively constant.”

For Cannon, equilibrium was a thermodynamic property of dead stuff. Two water baths connected by a pipe reach thermal equilibrium. A person in a sauna maintains an internal temperature of 37 degrees Celcius. Homeostasis is an active mechanism of regulation—a constant push to keep something constant. It requires millions of years of evolution and a constant intake of new resources to maintain.

Cannon’s concept of homeostasis would revolutionize the way we think about medicine. Doctors began to conceive many ailments and diseases as regulatory disorders. If a body becomes dysregulated, a doctor can prescribe a treatment to steer it back to homeostasis. One of the more striking examples in The Wisdom of the Body is an early description of how the body regulates blood sugar. Insulin was only discovered 10 years earlier, and Cannon describes its role in regulating blood sugar. Diabetes could thus be thought of as a disease of dysregulation. In 1920, children diagnosed with Type I diabetes would tragically die, often in as little as six months. Today, though still a very serious medical condition, Type I diabetes is not only treatable but can be managed with insulin pumps controlled by a cell phone.

I set out to write a talk that starts with Cannon and then thoroughly explores his influence, but I won’t be able to do that justice in a single talk. The pile of things I was hoping to engage with that have already been cut could be turned into talks themselves. Notably, homeostasis was a central motivating concept for the Cyberneticists and remained a tenet of systems thinking as cybernetics moved into management sciences. Dan Davies frequently writes about cybernetic homeostasis on his blog and in his fantastic book The Unaccountability Machine, and I want to engage more with Dan’s work here on argmin at some point soon.

But what can I do in 40 minutes? I’ve decided to present some results on how to model homeostatic systems as mathematical feedback systems. Some rather general principles have to apply if you want to keep a signal tightly regulated. Control theorists worked out these concepts in the 1950s, and I’m going to describe them as simply as I can, without any differential equations or Laplace Transforms.

What’s particularly interesting is these sorts of regulatory loops often adapt to unknown disturbances without a brain. Bacteria are very adaptive! I’ll cover some of the ideas from this fantastic survey by Mustafa Khammash, showing how adaptation occurs even in isolated chemical reaction networks.

Central to these homeostatic adaptations is the idea of integral control, and I will try to explain why some sort of integral control is necessary and emerges everywhere you look. And then I want to describe some of my own research in this area, which provides some general principles of using integral control to treat dysregulation in surprising ways. This hits on my favorite topic, individualized decision making, and hints at how we can improve ourselves by simultaneously being the treatment and control group.

I’ve been trying to get my thoughts on homeostasis in order for a while and decided to force my hand by submitting a talk abstract for a talk I haven’t written. I’m not sure if inventing deadlines is the best strategy to get things done, but it’s certainly a strategy. I’m fairly sure I’ll be able to make a compelling case, but these first talks always have a high degree of unquantifiable uncertainty associated with them. I’ll write more next week about how it goes.