How Do We Get Breasts Out Of Bayes Theorem?

[Epistemic status: I guess instincts clearly exist, so take this post more as an expression of confusion than as a claim that they don’t.]

Predictive processing isn’t necessarily blank-slatist. But its focus on building concepts out of attempts to generate/predict sense data poses a problem for theories of innate knowledge. PP is more comfortable with deviations from a blank slate that involve the rules of cognition than with those that involve the contents of cognition.

For example, the theory shouldn’t mind the existence of genes for IQ. If the brain works on Bayesian math, some brains might be able to do the calculations more effectively than others. It shouldn’t even mind claims like “girls are more emotional than boys” – that’s just a question of how different hormones affect the Bayesian weighting of logical vs. emotional input.

But evolutionary psychologists make claims like “Men have been evolutionarily programmed to like women with big breasts, because those are a sign of fertility.” Forget for a second whether this is politically correct, or cross-culturally replicable, or anything like that. From a neurological point of view, how could this possibly work?

In Clark’s version of PP, infants laboriously construct all their priors out of sensory evidence. Object permanence takes months. Sensory coordination – the belief that eg the auditory and visual streams describe the same world, so that the same object might be both visible and producing sound – is not assumed. Clark even flirts with the possibility that some really basic assumptions might be learned:

Plausibly, it is only because the world we encounter must be parsed for action and intervention that we encounter, in experience, a relatively unambiguous determinate world at all. Subtract the need for action and the broadly Bayesian framework can seem quite at odds with the phenomenal facts about conscious perceptual experience: our world, it might be said, does not look as if it is encoded in an intertwined set of probability density distributions. Instead, it looks unitary and, on a clear day, unambiguous…biological systems, as mentioned earlier, may be informed by a variety of learned or innate “hyperpriors” concerning the general nature of the world. One such hyperprior might be that the world is usually in one determinate state or another.

I realize he’s not coming out and saying that maybe babies see the world as a probability distribution over hypotheses and only gradually “figure out” that a determinate world is more pragmatic. But he’s sure coming closer to saying that than anybody else I know.

In any case, we work up from these sorts of deep hyperpriors to testing out new models and ideas. Presumably we eventually gain concepts like “breast” after a lot of trial-and-error in which we learn that they generate successful predictions about the sensory world.

In this model, the evolutionary psychological theory seems like a confusion of levels. How do our genes reach out and grab this particular high-level category in the brain, “breast”, to let us know that we’re programmed to find it attractive?

To a first approximation, all a gene does is code for a protein. How, exactly, do you design a protein that makes men find big-breasted women attractive? I mean, I can sort of imagine that if you know what neurons carry the concept of “breast”, you can sort of wire them up to whatever region of the hypothalamus handles sexual attraction, so that whenever you see breasts you feel attraction. But number one, are you sure there’s a specific set of neurons that carry the concept “breast”? And number two, how do you get those neurons (and no others) to express a certain gene?

And if you want to posit an entire complicated breast-locating system made up of hundreds of genes, remember that we only have about 20,000 genes total. Most of these are already involved in doing things like making the walls of lysosomes flexible enough or something really boring like that. Really it’s a miracle that a mere 20,000 genes can make a human at all. So how many of these precious resources do you want to take up constructing some kind of weird Rube-Goldbergesque breast-related brain circuit?

The only excuse I can think of for the evo psych perspective is that it obviously works sometimes. Animals do have instincts; it can’t be learning all the way down.

Sometimes when we really understand those instincts, they do look like weird Rube Goldberg contraptions made of brain circuits. The classic example is baby gulls demanding food from their mother. Adult gulls have a red dot on their beaks, and the baby bird algorithm seems to be “The first thing you see with a red dot is your mother; demand food from her.” Maybe “red dot” is primitive enough that it’s easier to specify genetically than “thing that looks like a mother bird”?

The clearest example I can think of where animals clearly have an instinctive understanding of a high level concept is sex/gender – a few gay humans and penguins aside, Nature seems pretty good at keeping its creatures heterosexual. But this is one of the rare cases where evolution might really want to devote some big fraction of the 20,000 genes it has to work with to building a Rube Goldberg circuit.

Also, maybe we shouldn’t set those few gender-nonconforming humans aside. Remember, autistic people have some kind of impairment in top-down prior-based processing relative to the bottom-up evidence-based kind, and they’re about eight times more likely to be trans than the general population. It sure looks like there’s some kind of process in which people have to infer their gender. And even though evolution seems to be shouting some really loud hints, maybe if you weigh streams of evidence in unusual ways you can end up somewhere unexpected. Evolution may be able to bias the process or control its downstream effects, but it doesn’t seem able to literally hard-code it.

Someone once asked me how to distinguish between good and bad evolutionary psychology. One heuristic might be to have a strong prior against any claim in which genes can just reach into the level of already-formed concepts and tweak them around, unless there’s a really strong reason for evolution to go through a lot of trouble to make it happen.