These videos are used in several of the presentations listed here:
Some were also used in my talk at this "Meeting of Minds" workshop in Paris 15-18 Sept 2007,
as part of the Eu CoSy project:
The videos below include demonstrations of different levels of competence in young children, and in some cases, surprising incompetence, which was later overcome, though I don't think anyone knows how the changes occur. There are also some non-human performers.
There are some conjectures in the presentations, especially the presentations concerned with mathematical development in young children, and its evolution.
See also the web pages on 'Toddler Theorems' here:
VIDEO 2: KITTEN CLIMBING DOWN A WIND-BLOWN TREE
Can any current robot do this sort of thing?
My ideas have recently been reorganised and extended by reading Piaget's last two books, and an important book by Annette Karmiloff-Smith
These exploration processes must be partly genetically driven in some of their abstract
properties, while details of both the exploration processes and the discoveries they lead
to are largely driven by details of the learner's environment.
Compare the notion of "parametric polymorphism" in Object Oriented Programming,
where a generic method performs actions and builds structures in ways that
depend on the types of entities provided as parameters when the method runs.
The notion of an exploration domain is closely related to the old AI notion of a "micro-world", to Karmiloff-Smith's notion of a "microdomain", and to the notion of a class (or related set of classes linked by methods) in Object Oriented Programming.
It is often assumed that what an intelligent animal or robot can learn must be closely constrained by its sensory-motor morphology and functionality.
This viewpoint ignores the role of the environment in presenting opportunities to learn and providing examples and evidence regarding what can happen in the environment, including processes initiated not by the learner but by wind, rain, gravity, other animals, etc. Learning about your environment is not the same thing as learning about your sensory motor signal patterns.
The sensory-motor analysis also ignores the similarities between types of process that can be performed using different morphologies: e.g. many carnivores use jaws to carry, manoeuvre and disassemble objects whereas primates (and some birds) use hands and/or feet.
The animals that can produce a-modal information structures using exosomatic ontologies (referring to things in the environment, not to sensory-motor signals) can learn things that have nothing to do with their own sensory or motor mechanisms, e.g. about effects of forces applied by one object to another, and effects of objects in constraining motions of other objects. They should also be far more flexible about transferring what they have learnt to new configurations where the sensory motor signals are very different, but what happens in the environment is very similar (e.g. a door is opened).
The comments on the videos relate to the notion of a "exploration domain" discussed in
Talk 90: Piaget (and collaborators) on Possibility and Necessity
And the relevance of/to AI/Robotics
A related presentation at Schloss Dagstuhl on 28th March 2011
And the discussions of domains in:
He achieves all his goals apart from joining two train components -- trying to join two rings together instead of a hook and a ring. His understanding of possible ways of linking two objects is clearly still under developed. (I was told a few weeks later that he had mastered the problems, though nobody had observed steps in the transition.)
Compare this with the intelligence of Betty the hook-making and hook-using crow, below.
She had no opportunity to learn anything about this from other crows or humans.
Betty made headlines world-wide in June 2002. See the video link to trial
7 here, highlighted on the left under "Movie":
To learn more, give to google:
betty crow hook 2002What the researchers did not point out in the original publications, but can easily be confirmed by looking at the other videos on the web site, is that Betty spontaneously made hooks from straight pieces of wire in at least five different ways. Even after successfully making a hook one way she notices and uses other possible strategies with no flailing about using trial and error. In each case she seems to know as soon as the action begins what she is going to do.
There are many idiotic comments online about Betty, including this one, which hints that acting naturally for one's species is incompatible with creative intelligence (e.g. using new materials to "act naturally").
"Crow that bent wire to retrieve food was acting naturally, scientists discover.
Oxford researchers initially thought Betty the New Caledonian crow had come up with idea on the spot but further study finds her breed adept at making tools"
I have argued elsewhere that such competences (and others) imply that many non-human species, and also pre-verbal children, must be able to use internal languages (forms of representation) that support structural variability, variations in complexity, compositional semantics, and use in making inferences.
The primacy of non-communicative language
Evolution of minds and languages. What evolved first and develops first in children: Languages for communicating, or languages for thinking (Generalised Languages: GLs)?
Those researchers are mostly interested in the fact that very young children and some other animals spontaneously try to help someone else.
A different sort of question that I am interested in concerns competences that must be prior to those motivations: what sort of cognitive mechanisms make it possible for a pre-verbal child to tell what a person is trying and failing to do, and to work out what could be done to provide assistance?
A child in one of the videos seems to be able to work out that the man carrying books needs to get the the door of the cabinet open. How does he or she infer that from the observed behaviour, probably never seen before? The child then adopts the adult's motive for himself (the altruistic step) and works out what to do to achieve the open door, then executes the plan.
Here the child seems to make use of a domain of possibilities that is not concerned with what he, the observer, can do, but with possibilities available to another agent. (I have called those "vicarious affordances" elsewhere: very important both for parents and for young learners.) What the child can do becomes relevant only after he has worked out what the adult wants to do, or needs to do.
What motivates the child to use those cognitive abilities to provide help is an additional question. Being able to think about motivation requires a domain of meta-semantic competence: being able to represent and reason about things that can represent and reason. That clearly starts to develop before a child can talk.
For both questions the answers are likely to involve a mixture of genetic mechanisms and competences learnt partly under the influence of the environment and partly under the influence of the genetic mechanisms, as explained in this introduction to the Meta-configured genome theory:The researchers, as far as I can tell, were not interested in the problem of explaining how the child is able to work out what he or she needs to do in order to provide useful assistance to the adult. That requires non-trivial reasoning abilities concerning spatial structures and processes and how they could relate to someone else's goal.
http://www.cs.bham.ac.uk/research/projects/cogaff/misc/meta-configured-genome.html (also pdf)
The learning mechanisms probably change dramatically during learning and development, since what is learnable changes. But this is not merely a matter of what prior content is available for each type of learning. Rather new ontologies, new forms of representation, new algorithms and new information-processing architectures seem to be involved.
Compare videos in which a parrot perches on one leg holding a walnut in the other foot, and alternating between holding it with foot and with beak rotates the walnut until it is in a good position for use of beak to crack it open.
(Compare: you are holding a cold saucepan by its handle using one hand. You wish to get into a position in which you are holding it by the rim of the pan using two hands, and with the handle pointing away from you. What actions do you need to perform to achieve that end state -- with or without being able to rest the saucepan on a flat table top during the process? What enables you, or the parrot, to work out a sequence of actions to achieve a new state?)
For further discussion of the issues raised by these
My slides for the 'Meeting of Minds' workshop in 2007, and the post-workshop notes on model-based semantics