Aaron Sloman
http://www.cs.bham.ac.uk/~axs
University of Birmingham, UK
This paper is one among many on this web site providing examples of how gaps in mathematical competences of AI systems, or, more generally, gaps in their spatial perception and reasoning abilities, are severe obstacles to explanatory power and practical usefulness of current AI systems. Future AI systems may bridge those gaps, but first we'll need a much deeper understanding of what is missing. This is one among a disorganised collection of documents illustrating those gaps, many of them concerned with abilities to reason about what is impossible, or necessarily the case.
Many forms of spatial intelligence lacking in current AI systems can be observed in pre-verbal human toddlers and other intelligent animals (including apes, squirrels, weaver birds, elephants, and many more).
BACKGROUND
There are growing numbers of impressive successes of artificial
intelligence and robotics, many of them summarised at
https://aitopics.org/, e.g.
https://aitopics.org/news.
Yet there remain huge chasms between artificial systems and forms of natural
intelligence in humans and other animals -- including weaver-birds, elephants,
squirrels, dolphins, orangutans, carnivorous mammals, and their prey.
(Sample weaver bird cognition here:
http://www.youtube.com/watch?v=6svAIgEnFvw.)
Don't expect any robot (even with soft hands and compliant joints) to be able to dress a two year old child (safely) in the near future, a task that requires understanding of both topology and deformable materials, among other things. (As illustrated in this video.) Enabling machines to understand why things work and don't work lags far behind abilities to perform tasks, often achieved by programming or training.
Much of everyday life involves complex mathematical structures, and (usually unconscious) mathematical reasoning.
The competences involved are related to abilities to perceive what is and is not possible in some physical configuration, which may be immediately present in your field of view, or remembered from past experience, or anticipated, or represented in a sketch, or merely imagined.
These possibilities have nothing to do with probabilities, except that thinking about what is probable presupposes abilities to consider possibilities and assign relative weightings to them. But considering the possibilities does not require any assignment of probability or likelihood.
What might mistakenly be construed as a special case of probabilistic inference is reasoning about what is impossible (cannot exist or occur) and reasoning about what is necessary (must be the case) in some situation. But impossibility and necessity have nothing to do with gradations of probability or likelihood, even though some thinkers confuse them with probabilities, i.e. 0% and 100% probability.
The poorly drawn figure below is merely intended as an aid to the understanding of a question about the consequences of attempting to put on a tightly fitting shirt or sweater by going through different processes.
Is it possible to put on the garment by inserting a hand into a cuff and pulling the sleeve up over the arm? Under what conditions could it succeed, or not succeed?
This requires a combination of topological and metrical reasoning: -- a type of mathematical child-minding theorem, not taught in schools but understood by some child-minders, even if they have never articulated the theorem and cannot articulate the reasons why it is true. Can you?
Merely pointing at past evidence showing that attempts to dress a child that way always fails does not explain why it is impossible. You can probably do better than that!
What sequence of movements could get the shirt onto the child if the shirt is made of material that is flexible but does not stretch much? Why would it be a mistake to start by pulling the cuff over the hand, or pushing the head through the neck-hole? What difference would it make if the material could be stretched arbitrarily without being permanently changed?
How many significantly different strategies are there?
How does the problem change if the sweater is much too small for you?
Can you turn a sock inside out by grasping it in two places, one with each hand, then moving your hands without letting go?
Is it possible to turn the sweater inside out by grasping it in two places, with left and right hands, then moving your hands without letting go?
If your answer is 'No' in either case, can you explain why it's impossible?
Could some of this relate to the types of reasoning carnivores require after a successful kill, before any meat is accessible?
This is one of several discussion pieces regarding vision and mathematical reasoning, pointing to serious inadequacies of current theories of vision in psychology and neuroscience, and inadequate visual systems in robots and AI systems. This is a part of the meta-morphogenesis project referenced below.
What sorts of mechanisms, or software extensions would have to be added to current robots, or theorem provers, to enable them to make these discoveries and reason about them?
Other examples include:
Meta-Morphogenesis project:
http://www.cs.bham.ac.uk/research/projects/cogaff/misc/meta-morphogenesis.html
Meta-Morphogenesis project overview.
This project aims at understanding the changes in types of information,
information processing, and information-using mechanisms, produced by biological
evolution, since the very earliest life forms, or proto-life forms existed on
this planet.
How can a physical universe produce mathematicians?
Another name for the project is "The self-informing universe".
http://www.cs.bham.ac.uk/research/projects/cogaff/misc/construction-kits.html
The important roles of evolved construction kits in biology
http://www.cs.bham.ac.uk/research/projects/cogaff/misc/midlog-talk.html
But the authors, not their theorem prover, translated the topological puzzle
into a logical form!
That report was dated 2011. A quick search revealed no follow up.
Note that there's a difference between being able to perform certain actions and being able to reason about possible and impossible alternatives to such actions.
Maintained by
Aaron Sloman
School of Computer Science
The University of Birmingham