Log in

No account? Create an account

Previous Entry | Next Entry

Tower of generate-and-test

In more than one of his books, Daniel Dennett talks about this concept, using it to categorize various kinds of creatures. This is someone's 2 page PDF diagram describing the ideas, but having just read one of Dennett's latest books, I'll give my own spin here.

* Darwinian creatures: essentially stimulus-response automata determined by their genes; what is generated is genotypes, what is tested is phenotypes reproducing in the environment.

* Skinnerian creatures: use associationism or reinforcement learning (called conditioning in lab experiments); what is generated is (initially) random behavior, what is tested is the consequences. Darwinian constraints hopefully keep the creature from killing itself outright before it has learned; with more altricial animals, parents may 'teach' their children by setting up the right environments for their children to play in.

* Popperian creatures: have models of the world good enough to generate hypotheses or imagined behaviors and evaluate them by their predicted consequences. There's a small cottage industry of experiments trying to show whether corvids or apes have insight in their tool use, vs. just using a past library of Skinnerian fooling around. Complicated by, well, how does one get such models? Almost certainly by a childhood or lifetime of Skinnerian experimentation, with Darwinian biases (like expecting object persistence) helping: think of a baby trying to stick things in holes, until eventually learning not just that round holes won't go in square pegs, but 'why'.

* Gregorian creatures: this is where the standard description, as captured on the PDF, loses me. But to me it seems that the next level is obviously language, or more generally, imitation, both of which seem uniquely human[1]. To try to stay on theme, what is generated and tested is other people's behavior, which we can then learn from.

In looking for summaries, I saw mentions of Dennett in some older book proposing a fifth level, of the scientific method. At first I dismissed that as just being a mash of Skinnerian experimentation and Popperian modeling. But note that each successive level 'learns' orders of magnitude faster than earlier levels. Likewise, counting from 1650 or later, the scientific community has learned the world far more broadly and deeply than all the millennia of human civilization before it. I feel it's too simple to say that's just due to scientific method: tools like lenses, good clocks, and the printing press helped a lot. But still, there's a big change.

Some other analogies come to mind. 'Gregorian' learning is like enhanced Skinner: instead of being limited to trying your own behaviors at random, you can learn from other people's behavior. "Did you hear what happened when Bob ate that mushroom?" Gregorian, or at any rate human, thinkers are all about giving reasons for our behavior and inferring reasons for other behavior; science sort of enhances Popperian models with reasons, going from "I have observed predictable regularity" to "I can explain predictable regularity and make further predictions from deeper principles."

See the comment count unavailable DW comments at https://mindstalk.dreamwidth.org/502273.html#comments


Damien Sullivan

Latest Month

June 2019


Powered by LiveJournal.com
Designed by Lilia Ahner