Pallium

Telling look-alikes apart

New: a region with random wiring that spreads the look of a bush over many cells, few of them active at a time. Lessons from pain are learned on those cells.

Result: the animals can now learn poison that only a combination of features gives away. A lesson stays on the poison and no longer spreads to the good look-alikes.

Since the association stage the animal learns from pain which bushes to avoid, on the look of a bush: a handful of colour features. That works when poison and good food differ in some feature. It fails when what makes a bush poison is a combination: here a bush is poison when its two colour features are both high or both low, and good when one is high and the other low. Whatever a lesson on the features teaches about one poison, it teaches the opposite about the other.

colour feature 1 colour feature 2 ☠ ☠ No single feature tells them apart Poison is low-low or high-high; food is mixed. A lesson after the low-low poison makes "low" bad: the good low-high bush looks bad as well, and the high-high poison looks safer than before. The pallium gives each corner cells of its own, so a lesson lands on one corner only.
The four kinds of bush in every life (two colour directions are drawn anew per life). Red with a skull: poison; green: food.

This stage adds a new region in front of the learning, the pallium.

The circuit

  • pallium: 128 cells. Each reads the colour features of the eye that points straight ahead, through synapses of random strength. The wiring is random and evolution does not shape it: what matters is that every cell is different, not what each one is.
  • Few cells at a time. Only the two cells that receive the most input fire; the others are held silent. (Real brains do this with inhibitory cells that every active cell excites; the model takes the shortcut and keeps the strongest two, called k-winners-take-all.) Because so few cells fire, two kinds of bush that share most of their look share almost none of their active cells.
  • The lesson. Pain teaches the synapses from the pallium onto the bad cells, by the same rule that has taught the synapses from the look since the association stage. The old route stays and still learns too; evolution sets how fast each route learns.

Only the eye ahead feeds the pallium: at a bite, the side eyes see neighbouring bushes, and those would be blamed too. An animal turns to a bush before it bites it, so the eye ahead sees the one that counts.

Biology

The pallium is the roof of the forebrain of all vertebrates; in mammals most of it became the cerebral cortex. Its oldest part handles smell: in the piriform cortex every cell receives input from a random handful of smell channels, and only a few percent of the cells respond to any one smell. Insects arrived at the same design on their own in the mushroom body: about fifty smell channels fan out onto thousands of cells, a few percent of which fire, held down by a single inhibitory cell that all of them excite. Such a random, sparse expansion separates inputs that overlap, which is why a fly can learn to avoid one smell and not its look-alike.

A fish turns to its prey before it strikes, and many fish have a patch of the retina with denser cells for this; the eye ahead stands for that. That the first vertebrate pallium did this job for vision is a guess of the model; that it began as such a sensory region is not.

The world

The world of the affect stage, with two changes. The four new kinds of bush in every life have looks on the corners of a square, as in the figure above. And the patches are mixed: every patch holds every kind, so where a bush grows says nothing about what it is, and the look has to do the job.

Did it work?

The pallium is used in every seed. Switching off all learning costs more than the pallium alone, because the look still learns as well. Against the old brain, the new brain is ahead in every seed when the two share a world. Measured apart, one seed is level: one life per animal makes that measure noisy.

In the lab

The sharpest test is outside the world. One animal at a time eats from four new bushes, two poison and two good, twice each, and is then shown each bush again. How likely is it to bite?

Learning on the look alone makes the animal shy of all four bushes alike: it cannot tell the kinds apart. Learning through the pallium alone more than halves the bites at poison and leaves the good bushes almost untouched. With both routes learning, as in the world, the animal is shy of the good bushes too, but less than of the poison.

Full results

Of the states of the affect stage only the one that settles the animal at food is still used, a little; fleeing and being fed no longer matter. In mixed patches the place says little about a bush; its look says more.

  • Mixed patches. In patches of mostly one kind, the place answered the question first: a poison bite or two and the alarm carried the animal off, and knowing a poison by its look saved only the first bite at the next bush of that kind. Evolution then turned the lessons down.
  • Fewer cells at a time. With four cells active instead of two, a lesson spread twice as far to the good bushes.
  • Evolution still damps aversion close up. In every population the inputs that see a bush right ahead hold the bad cells down: approach what is in front of you. A lesson has to overcome that before it can stop a bite.

From this stage on, how strongly the look inputs adapt (the habituation of chapter 1) is set by evolution too.

Runs