Neuroscientists suggest Hebbian learning. Better known as "neurons that fire together, wire together" which can be roughly rephrased to "neurons identify correlations in the input". They do this in a quasi-layered environment differentially modulated by waves of the materials used to alter their connections (a.k.a. trophic factor).

I believe it's like a tree that's sprouting.

Imagine the sensory pathway from sense organ to brain. As information travels it preferentially goes in one direction as opposed to another. When it reaches a dead end it splits and forms its own branch. Then the next bit of information will reach this new dead end, and split again. And so on until everything is processed and categorized. Basically a graph search problem in a self-modifying graph where you can have cycles and feedback loops and layers and many other structures that are not yet captured by the state of the art in ML, and where each node is specialized for a specific task.

Somewhat Analogous to tokenization?