Learning to Learn in the Development of Action. In
Action As An Organizer of Learning and Development: Volume 33 in
Volume 33 in the Minnesota Symposium on Child Psychology
Edited by John J. Rieser, Jeffrey J. Lockman, Charles A. Nelson
I had never read it, though I heard her talk about the work at
the CoSy Project workshop in Sept 2007.
She addresses the question whether great variability in some aspect of the environment can *prevent* associative learning from happening:
"Variable contents may help discourage simple associative pairings."
I think is somewhat misleading.
It's not a case of discouraging, but simply not enabling the appropriate frequencies to be encountered. If variable contents are encountered often enough the associatiative learning *will* happen and that's how fluency in complex new areas (including language learning, mathematics, music, athletic performances, etc.) develops.
Each multiplication of two numbers is different from every other, and the first time you need to know what 5 x 13 is you may have to work it out. But if you encounter that problem often enough you can 'cache' the result. (Some people only need to do it once.)
Humans transfer vast amounts of knowledge from temporary results of reasoning to long term associative stores, in some cases only if the results are repeated. Presumably that can also happen to a subset of the arboreal locomotion problems. (There is also some one-shot learning, e.g. listening to a lecture on a mathematical proof and remembering how the proof goes, or Mozart hearing a performance and then writing it down).
But when the combinatorics imply that the total set of possibilities for action in a type of situation is very large, many that are encountered for the first time will require the use of reasoning.
So instead of talking about variety 'preventing' learning, we should talk about what doesn't happen as preventing the learning).
It's the distinction between dealing with novel situations 'online' (the word she uses) and dealing with them 'offline', e.g. during planning of possible actions, reasoning about what would have happened if, making predictions about consequences of something observed, creating explanations of observed evidence.
Adolph's examples all (or nearly all?) involve dealing 'online' with novelty, and she emphasises that explicitly.
This means that the performance of a complex action is done incrementally and each sub-step is actually performed, then a subsequent sub-step selected from the possibilities that then become available (sometimes from a continous range of possibilities rather than a discrete set).
This does not require the information processing system to be able to represent multiple branching sets of future states and actions. There is only one set of branches at any time, the set of possible next things to do -- where the agent is already physically poised to select and do one of them.
The situations we were envisaging testing were computationally more sophisticated because they required envisaging choices in possible *future* situations, where the choices could lead to new choices.
Perhaps 'anticpatory' could be contrasted with 'online' in this context.
I don't think I have ever read, or written, or said anything very clear about the difference between online creativity and anticipatory creativity, though I point to the difference in the video of the child pushing a broom, which I think you have seen me use: http://www.cs.bham.ac.uk/research/projects/cosy/conferences/mofm-paris-07/sloman/vid/ In that example we can distinguish the two sorts of creativity.
Online creativity involves dealing with situations where the broom handle is caught between rails, or when the broom has been pushed into a skirting board that restricts its motion, and maybe when the broom is being steered down a corridor for the first time.
Anticipatory creativity is demonstrated when in advance of a doorway leading off to the right the toddler starts moving the broom handle to the left so that it can then be pushed forward into the room when he has reached the doorway. That's only one step anticipation, whereas I think the orangutans have to solve more complex problems.
In part they can do the more complex anticipation by ignoring details that cannot be discarded in the online case, e.g. the precise positions of limbs, the precise contents of the visual field of view, the precise momentum produced by previous movements, etc.
Making the problem more abstract by ignoring details increases the sophistication of the mechanisms and forms of representation required (the ability both to derive abstractions from detailed percepts and detailed sets of motor signals, and also to use the abstractions in producing detailed behaviours), but that greater sophistication results in much reduced computational load -- there's much less information to process in selecting the actions.
It seems to require a large extension of the brain to process that kind of much-reduced information! (Including frontal lobes?)
Is what I have just written familiar from any of the animal behaviour literature? I don't recall ever having read it in the context of AI/Robotics, but it is so obvious that someone must have noted it.
School of Computer Science
The University of Birmingham