Bayesian Brain Slugs

Last week a friend, Joe, was visiting from Dunedin, bringing a crop of stories from the ten years of study he just finished. Here's my garbled version of one of those stories, because I want to remember it, and if I try to remember it in a month's time all I'll come up with is a handful of disconnected assertions. It is a story of the origin of the Bayesian brain.

The earliest nervous systems began in (what Joe calls) roomba slugs: animals three cell-layers thick, able to sense only what they lie upon, navigating a practically featureless, practically endless layer of edible algae on the seafloor. (Effectively, at this point, their bodies are one-to-one maps, and the map is all the sense-data they've got). All they could do was move randomly and eat what they stood on. Each direction of movement was almost as good as any other - but 'almost' becomes a sentence of life or death, in evolutionary time. The least comparative inefficiency ends the less efficient lineage. So rudimentary systems developed to guide movement toward food, away from sparse patches, etc. Higher concentrations of glutamate in the water tended to indicate algae in that direction. (Because glutamate is a protein component released by bacteria associated with the algae? I think? I've lost the details on that part). The slugs didn't require a nervous system to do this, because the legs didn't have to communicate with each other to be effective: if each leg was rigged to pull harder in the presence of more glutamate, the net effect would be to automatically move the slug towards the source of the glutamate (although the legs on opposite sides of its body would be constantly tugging against one another, so it's not ideal). Glutamate is the most common excitatory neurotransmitter in vertebrates, and that may be because it used to be the direct 'go' signal to slug legs, and that's the mechanism onto which everything else was later bolted. (Why just vertebrates, since invertebrates also descend from roomba slugs? I do not know).

Things changed at the point where slugs began to eat each other. (There is a more complicated story to be told about when that point was, how slug population and algal density regulated each other, etc). Hitherto, slugs had not had a reason to distinguish between objects. There weren't any objects. If you bumped into another slug, the most you needed to be able to do was bounce off it and move in another direction. (Presumably that's a half-truth? Because the presence of other slugs would indicate competition for algae, so you'd want to be evenly distributed. Ignoring that for now). But once some kind of modification meant that one slug, having crawled on top of another, could begin to digest it, it became very advantageous to slugs to have a method of detecting other slugs, and to hook that up to a type of cell which emitted glutamate and told the relevant slug legs "Go! Go now!"

The rudimentary form of this is a chemical signal which says, "Slug cells have broken!" which means either "I am being eaten!" or "I am eating something!" (Humans still use this system. I've forgotten the name of the chemical, but when it is released in our bodies it is a 'wound' signal, unless it's in our stomach, in which case it is a 'you are digesting' signal). Knowing that you are being eaten is good, but not the best possible way to avoid being eaten. Detection at a distance? Sight isn't good in murky water, the kind of basic light-sensitivity you can get right off the bat only kicks in when the other slug is very close. Sense of touch likewise, and as a slug, you aren't complicated enough to make a water-tremor-detection lateral line system like what fish have. Sensing chemicals is also not excellent, because chemicals persist in the absence of their source, and by the time the concentration gradients get useful, the other slug is already closer than you'd like. (That is something else I want to know more about, because it seems to tell against the glutamate-sensing slug-legs thing. How hard is it to evolve a really useful nose?)

But at the moment, nothing non-living in the ocean has an electric field. And since proteins already in your body are affected by electric fields, and you already have some internal processes which use that, getting electroreceptors set up to say 'Another slug incoming' is a plausible thing for evolution to do. At which point an arms race begins: predator and prey (larger and smaller slugs, at this stage) are each advantaged by being able to detect the other before the other can detect them, so as to begin chasing or fleeing at the right moment.

(The prey-slug does not, funnily enough, want to flee early enough that it definitely won't be eaten. That would be inefficient. Fleeing costs energy. There is an optimal moment where 'run away' becomes better than 'stay still and keep eating' in terms of acquiring energy and not spending more of it than necessary. That moment is when the slug has about a forty-nine percent chance of being caught and eaten).

The problem with electroreception is that water has a weak background electrical field. At a certain point, making your electroreceptors more sensitive doesn't let you sense other slugs further away, it drowns you in background static instead. This is the moment when the slugs have to start doing Bayesian inference.* If your electroreceptors are so sensitive that they always fire sometimes, you need a system to determine what pattern of firing means, 'It is more energy-efficient to assume that a predatory slug is approaching me than to sit here and keep on eating.' (This is analogous to what human ears do: ears can accurately detect sounds whose vibrations have considerably less energy than the random motion of air molecules). It doesn't take a very complex network to start doing statistics like this: the electroreceptor, and then a layer of neuron-esque things the firing of each of which represents a certain hypothetical position of the incoming slug, and if the neuron-esque things are arranged in order of distance, a cluster of strong signals will say, 'the slug is probably somewhere around distance Y' and some particular cluster of strong signals will say, “It is x% likely that a slug is approaching you and has just reached the optimal run away distance'. (The details of this system have left my mind very quickly, which suggests I never quite understood it, although Joe did draw diagrams).


And after that, to make a long story short, you get the Cambrian Explosion. (Or the Cambrian Radiation, if you want to be peaceful about it). Nervous systems really deform the efficiency landscape. All of a sudden it has all kinds of valleys of optimisation to sit in, when before there was really only the 'Best slug' valley. You can be best attack slug or defence slug or hit-it-with-a-limb slug or run-away-faster slug or hey-I-can-float slug. Once you have a system which can make predictions and coordinate responses, it's quite easy to apply different inputs and attach different behaviours to the outputs, and your efficiency landscape acquires too many dimensions to put in a graph. The Radiation/Explosion is defined by the appearance of burrows in the fossil record, because they're good evidence for nervous systems: it isn't easy to burrow unless you meant to.

That's the linear bit of the story. Other bits:

People (says Joe) often talk about the brain as though it was muddled together any old how. This is analogous to Richard Dawkins pointing at the vulnerable nerve in the neck of the giraffe and saying, “Look what foolish results evolution can produce: how unlike something designed.” But again and again in evolution, you see features sitting right at the local minimum of least energy expenditure. (I don't know if Joe meant that the giraffe nerve serves a function in being so long, or if that was only an illustration and giraffe necks really are ridiculous). But something apparently non-beneficial may, like the running-away distance which lets only fifty-one percent of slugs survive, have less-obvious advantages. And the brain, which costs a heck of a lot of energy both to use and to maintain, has nevertheless developed as efficiently as possible. It has no central coordinator saying, “We have almost used up our glucose supply, slow down you neurons.” Every neuron, on its own, has to be as efficient as possible, or the brain overall will exhaust its glucose supply. (This is what happens in a long epileptic seizure). Computers are much better at some things than brains, but they are much, much less efficient, because brains evolved in order to get better energy-conservation results than the next slug over.

People also talk about evolution as a heroic climb, but it acts more like a series of tumbles: human cognition seems to have developed like a rock being rolled down an optimisation curve. Once it starts rolling, and humans get squishier but smarter, there's no good way to climb back up towards being less squishy and less smart.** Intelligence creates a landscape of problems which intelligence is good at solving. (This also applies to squid, although admittedly not as much).

The idea of the Bayesian brain is: human cognition doesn't so much combine incoming sense data into a direct picture of the world, as produce a stable hallucination of what our internal model tells us should be in the world, and then use sense data to make corrections. Instead of going, “I detect lines and circles: those add up to wheels/frame/handlebars: we detect all of wheels, frames, and handlebars: ding, motorbike!” from the bottom up, the brain goes, “The context makes me expect a motorbike. Does anything we perceive indicate the absence of a motorbike? Well, do any of the lines and circles we see indicate the absence of wheels/frame/handlebars? No? Well, I'll go ahead and perceive a motorbike then.” At every level of perceptual analysis there's a little loop checking what is perceived against what is expected – which complicates the idea of a distinction between perception and cognition, and which explains a number of optical illusions, among other things.

In fact it sounds rather as though it might Explain Everything, except Joe has stated his scepticism for any theory which discovers something nifty and then says, “This single nifty thing Explains Everything!” (The actual research he's drawing on mostly, I think, restricts itself to really sturdy analysis of four-layer neural systems in the human ear).

And that is pretty much all I can remember. I will post it, and Joe can tell me how much it's wrong and send me citations for the other bits. (Well, how much he thinks it's wrong. There may be other neural evolution cults than the Bayesian Brain Slug Cult? But it sounds like a good one to me).





*The main equation of Bayesian inference basically says 'How likely would the outcome I observe be, given my hypothesis about what caused it?' For a while I understood the equation itself, but in a wibbly, in-and-out way which didn't linger.

**Although one of the things Joe suggested was that sponges, which don't have nervous systems, actually descend from things that do, used the nervous systems to get through a difficult hump of slug competition, and then lost them again later. That's highly speculative. But I do love the way each successively-more-complex marine phylum, going through sponges to jellyfish to worms to different worms to crustaceans, has some species which say, "Yes, I know I could be more mobile/intelligent/complicated than that simpler phylum, but you know what I really want to do? Sit absolutely still and filter-feed. Filter-feeding is great."