Why we seek out pleasure: the Symmetry Theory of Homeostatic Regulation

Why do we seek out pleasure- what Freud called the “pleasure principle“?

More accurately: why do we seem to seek out pleasure most of the time, but occasionally seem indifferent to it or even averse to it, e.g. in conditions such as anhedonia & depression?

My answer in a nutshell:

  1. Our brain networks are calibrated to the environment such that their symmetry gradients are tied to survival requirements. This is the core algorithm by which our brains regulate homeostasis. (Argued below)
  2. Symmetry in the mathematical representation of phenomenology corresponds to pleasure. (Argued in Principia Qualia)
  3. In combination, then, pleasure-seeking is what it feels like when our brain follows its core (ancestral/default) algorithm for maintaining homeostasis.

An algorithmic model of brain-dynamics-as-symmetry-gradient-climbing across multiple scales:

This is a simple model of a complex system, and as such it can only be approximately correct. But I think it’s also substantially correct, and that any future gears-level model will need to be compatible with this model. It centers around coalitions & symmetry & attractor basins.

The basics: the brain is a complex chaotic system centered around self-organized criticality. This criticality is based around coalitions of neurons, which “vote” (and thus produce activity cascades) with their firing patterns.

The brain has many attractor basins, but the central & most important one is harmony (and harmony ultimately boils down to symmetry). Thus the brain is always searching for actions, plans, and environments which lead to activity cascades toward more harmonious/symmetrical states. But never too harmonious! We can think of boredom as a very sophisticated ‘anti-wireheading’ technology that gradually pushes our networks’ symmetry (and thus stability / disposition toward inaction) down when our environment becomes too predictable. Our brains are calibrated to dynamically circle around the symmetry basin without ever falling in and getting stuck, since the computationally-useful part of this is the high-dimensional symmetry gradient which is finely calibrated to our environment and homeostatic requirements… not the actual symmetry itself. (You’re a cruel one, Azathoth…)

Basically, brains try to climb their internal symmetry gradients through both internal & external means: when we move our bodies, change our environment, make plans, listen to music, eat food, respond to threats, and so on, we’re increasing our internal symmetry gradient (or at least preventing it from dropping). Symmetry gradients are how our brain internally represents value, and much of the brain’s complexity has to do with calibrating these gradients to the environment in order to make them computationally useful. I.e., to make the contextually correct things increase & decrease symmetry in core regulatory networks.

Getting a little more technical:

A good algorithmic theory of cognition will collect, unify, and simplify a lot of things that look like odd psychological quirks, and recast them as deeply intertwined with, and naturally arising out of, how our brains process information. I’m optimistic that Symmetry Theory will be able to do just this- e.g.,

  • Cognitive dissonance happens when two (or more) patterns in your head are battling for your neural real estate, and they’re incompatible– i.e., they’re collectively dissonant/asymmetrical.
  • Denial is what happens when your brain attempts to isolate/quarantine such patterns, and is actively working to prevent this tug-of-war for neurons.

This model implies that your brain can evaluate the “internal consistency/harmony” of a neural pattern, and reject it if there’s a negative result- and also the “simulated relative compatibility/harmony” of two neural patterns, and try to keep them isolated if there’s a negative result. I’d suggest the best way to understand this is in terms of projective geometry, resonance, and symmetry: i.e., to evaluate a pattern’s “internal harmony” and whether it ‘runs well (is stable) on existing hardware’, the brain uses principles of resonance to apply certain geometric projections (high-dimensional-to-lower-dimensional transformations) to the pattern to see if the result is stable (unchanged, or predictably oscillating, or still strongly resonant) under these transforms. Stable patterns are allocated territory; unstable ones (=dangerous neural code) are not. The internal mechanics of this will vary across brain areas (based on the specific resonance profile of each area) and emotional states, which might contribute to how certain types of information tend to end up in certain brain regions. Likewise, this could explain how moods coordinate information processing– by changing the resonance landscape in the brain, thus preferentially selecting for certain classes of patterns over others. A core implication of this model is that different kinds of dissonance will drive different kinds of behavior (feel like different kinds of imperatives), and based on what action is needed, a mood may create (or be the creation of) a certain kind of dissonance.

Stepping back:

Ultimately, this is the first move in an attempt to subsume all of cognitive processing under the mantle of symmetry (similar to Smolensky and Legendre’s Harmonic Mind hypothesis, but not limited to phonology), just like QRI is subsuming all valence under the mantle of symmetry. I think this is not merely ‘a useful way to model the brain’- I think it’s literally how the brain works. We could call this the Symmetry Theory of Affective Cognition – with the added implication that most things the brain does, from high-level cognition to low-level autonomic regulation, are all variations on this core theme. More technically, I’d suggest calling this the Symmetry Theory of Homeostatic Regulation.

This package of mechanisms doesn’t seem like the only way to build intelligent systems, but it does seem like a particularly good one that can start simple yet climb the ladder of abstraction, allows implicit social coordination from the bottom-up (“‘vibes’ are like, a real thing, man”), and can evolve in incrementally-useful forms. Symmetry is just an insanely powerful principle to build or evolve a computational system around.

Waxing philosophical, I think this is my favorite high-compression description of what humans fundamentally are— the essence of this “human condition” thing we seem to only have indirect access to. I.e., to be human is to be a complex dynamic physical system which maintains homeostasis via a strong symmetry attractor, on one hand, yet on the other hand also has sophisticated anti-wireheading technologies that ensure we never stay deep in this attractor for very long. If we want to build a computer with human-like qualia, human-like drives, and human-like cognition, we’d need to build it around these principles. (Necessary, but not sufficient, etc)

Implications & extensions:

This is only a down payment on an actual gears-level account of cognitive processing. A few brief suggestions: it’ll probably be fruitful to:

  1. try to match natural failure modes of this system to various psychological & psychiatric conditions- e.g., let’s page through the DSM-V and see if we can spot dysfunctions which correspond to breakdowns in the processes I talk about above, and if that offers a new way to carve the problem space of psychiatry;
  2. dig into the “brain regions as neuroacoustic chambers” angle;
  3. explore how the brain’s symmetry gradients are initially (developmentally) correlated with the environment, and how the brain actively & passively maintains this correlation (Adam Safron is doing interesting work here);
  4. see if this sort of system could emerge from a handful of simple, known neural rules (e.g., “neurons that fire together, wire together”);
  5. try to rebuild various top-down psychological theories (like CT, or CFAR stuff) in terms of Symmetry Theory;
  6. try to reclassify primary neurotransmitters such as opioids/dopamine/serotonin in terms of their local & global effects on neuroacoustics. (Andres Gomez Emilsson is doing interesting work here, and will be speaking about it on June 7th.)

In particular, I think it’d be useful to brainstorm about Symmetry Theory, affective disorders, and effective treatments. E.g., listening to music might be unexpectedly effective at helping regulate mood.

Looking more abstractly, we could also view social dynamics, moods, and memetic desire through the lens of transmissible neuroacoustics (Girard’s mimesis).

Symmetry Theory’s relationship to similar theories:

How does this framework compare to e.g.,

  • Schmidhuber’s Compression Drive, which posits that the brain-as-learning-organ is constantly seeking efficient compression schemes for input and storage, and actively seeks out elegance to improve its compression library;
  • Friston’s Free Energy Principle, which argues that “[A]ny self-organizing system that is at equilibrium with its environment must minimize its free energy. The principle is essentially a mathematical formulation of how adaptive systems (that is, biological agents, like animals or brains) resist a natural tendency to disorder. … In short, the long-term (distal) imperative — of maintaining states within physiological bounds — translates into a short-term (proximal) avoidance of surprise.”
  • Seth et al.’s predictive coding / Bayesian Brain framework, which suggests that cognition revolves around sophisticated predictive models of sensory input, and will reduce bandwidth requirements by only passing surprises up the cognitive hierarchy.

In short, all these theories seem to be like different blind men examining different parts of the same elephant: they’re essentially doing the same thing, just in different ways. Here’s how I see the Symmetry Theory of Homeostatic Regulation in this context:

  • Mechanism: Schmidhuber et al.’s work operate on Marr’s computational level, and describe the high-level logic of the brain, whereas Symmetry Theory is an algorithmic description and supplies a mechanism by which theories like Compression Drive could be implemented. This focus on mechanism lets us see more details, e.g. how patterns can compete, how patterns can interact with substrate & selection landscapes, how boredom/denial/etc work, how pattern compression occurs, etc. (There are probably multiple algorithms you could build CD out of, but ST seems by far the best one.)
  • Generality/scope: symmetry gradients seem usable for regulation of homeostasis at all levels, from simple autonomic functions to complex cognition, whereas compression drive deals with information, and as such seems narrower in scope & requires more complexity/context. Predictive coding theories may fall somewhere in the middle, and essentially describe efficient information flows in a complex ‘symmetry attractor type’ system.
  • Frame-invariance (ease of application): One of the biggest problems with the class of theories Schmidhuber’s work belongs to (Free Energy, Bayesian Brain, etc) is that it’s ambiguous how they actually apply to any given system. Symmetry Theory should be a lot easier to apply since symmetry is literally the thing that enables frame-invariance. (This is not meant as a critique of Schmidhuber’s/Friston’s/Seth’s work, of which I am a huge fan.)

In short, I think symmetry, and the Symmetry Theory of Homeostatic Regulation, can play a part in unifying & contextualizing theories in this class, since the core of each of these theories is some different feature or flavor of symmetry/regularity. However, we haven’t done the actual work of unification yet, and need to do some deep thinking about what kinds of problems each approach is best at, and why.

Question for my readers:

What would you want out of a theory like this? If you’re the customer, what’s the product you’d like to see? How can I translate this from a bunch of words into something that makes your life better?

Inspirations & surrounding literature:

Principia Qualia; many conversations with Andres Gomez EmilssonSmolensky‘s Harmony theory of neuro-linguistic processing; Predictive coding / Bayesian Brain work by FristonSethSchmidhuber, and others; conversations with Adam Safron, and his entrainment model of orgasm; Perceptual Control Theory; Marblestone et al.‘s three hypotheses; David Pearce’s notion of the pleasure-pain axis as computationally relevant; Lin & Tegmark‘s findings of symmetry in deep learning; Trivers‘ arguments about symmetry & developmental biology; various conversations with PG & Romeo & Randal & Radhika, & probably others I’m forgetting.