NOTE Added 16 Jan 2006:
Some of the ideas behind this web site are explained more clearly with examples in this web page on recombinable orthogonal competences.
The points are made in relation to a distinction between learning about two kinds of causation, Humean and Kantian in this presentation (on two views of a child as a scientist).
Related points are made in a partly overlapping presentation on A (possibly) new theory of vision.
Introduction: understanding structure, motion and causality I am discussing with several people the possibility of a project that will be closely related to, but complementary to, CoSy. It is especially close to the PlayMate scenario in CoSy.
The new project will focus on trying to understand some of the exploration-driven learning mechanisms found in many animals, with particular emphasis on the development of understanding of spatial structure, motion and causality, as opposed to, for example, recognition, classification, prediction or tracking of objects.
The project is not primarily concerned with the probabilistic (Humean) notion of causality that is currently much discussed in AI and child psychology, which is concerned with correlations and how they are tested (as illustrated by Alison Gopnik's invited talk on children as scientists at IJCAI05, and in the last few papers available from her home page).
Another recent example of research in that field is Mark Steyvers, Joshua B. Tenenbaum, Eric-Jan Wagenmakers, Ben Blum, 'Inferring causal networks from observations and interventions' Cognitive Science 27 (2003) pp 453-489 http://web.mit.edu/cocosci/Papers/steyvers-etal-2003.pdf
In contrast with such work, our investigation of causation will focus mainly on learning to perceive, understand and deploy the deterministic causal relations involved in change of structure or change of relations between structures. Kant argued in his Critique of Pure Reason (1780) that Hume was wrong and that we have a conception of causation that goes beyond mere correlation and degrees of probability, but is deterministic. I believe Kant was right and that the ability to understand this notion of causation was a biological precursor to the development of mathematical capabilities. But exactly what the Kantian thesis is remains problematic: neither he nor Hume had conceptual tools for thinking about the design of a working system that learns about causation of either kind.
This is not to deny the importance of Humean concepts of causation: they are used either when we discover causal links that we do not understand (e.g. a child learning that flipping a switch can make a light go on or off), or when there are multiple causes for a particular event, e.g. age, strong winds and gravity combining to make a tree collapse. But the history of science shows that often what is initially thought of as Humean causation is later seen to have been ill-understood Kantian causation, where an underlying deterministic mechanism explains the observed correlations (e.g. fire producing smoke and ash from a pile of wood, or eating certain plants making people die).
It is sometimes argued that quantum mechanics shows that there is a deeper notion of probabilistic causation underlying the deterministic causation of physics, chemistry, biology, etc. But that is not an issue discussed here. (There is a kind of deterministic relationship between structure and long-run frequency that often underlies what is normally thought of as probabilistic causation, based on the mathematics of combinations and permutations, but that's a topic for another discussion.)
In contrast, if the wheels are linked by some mechanism hidden in a box
I discussed our ability to understand and reason about spatial change using spatial forms of representation in http://www.cs.bham.ac.uk/research/cogaff/sloman-analogical-1971/, a paper presented to IJCAI in 1971, criticising the logicist philosophy of AI presented by McCarthy and Hayes, in 1969.
The list above presents just a tiny sample among thousands of facts about structure, motion and causality that every normal human child is capable of learning. Some of them are also learnt by some other animals. It is likely that nest-building birds and tree-climbing animals have to learn many examples that most human children do not normally learn, though they may be capable of learning some of them. Animals that fly may learn many things that humans never learn, e.g. about the fine-grained control of motion in flight, though some humans learn the underlying aerodynamical principles that the flying animals never learn!
Nothing is claimed at present about how such things are learnt. It is possible that in every case a child first learns the causal connections as mere correlations (Humean causation) and then later comes to understand the deterministic (Kantian) causal explanation of what is happening. Alternatively it may be that in some cases Kantian understanding can be immediately extended to new cases without having to go through an intermediate Humean understanding of those cases, e.g. seeing for the first time that a rigid circular plate (e.g. lid of a coffee tin) can be used to perform some of the same functions as a long thin lever, or perhaps seeing for the first time that bending a piece of wire to form a hook can enable you to make an object move towards you, as Betty the New Caledonian Crow did in Oxford, much to the surprise of researchers there in 2002.
Learning such things, or being given the information innately, will be required in a robot that can manipulate objects intelligently, for example in order to help a disabled person with everyday tasks in a kitchen.
However, fast, fluent, unintelligent manipulation is possible without understanding. One kind of ability, the ability to behave expertly, may be implemented in appropriate kinds of dynamical control systems, whereas the other kind, the ability to reason about what behaviour is possible and what its consequences will be is what we are concerned with here, and that does not necessarily go hand-in-hand with speed and fluency.
One reason why causation (usually, though not always deterministic) in virtual machines that we have built and understand well is important is that it at last helps us understand much older examples of causation in virtual machines, namely mental causation, which has been very puzzling to philosophers had scientists for a long time. For more on this see this online presentation (PDF).
I mention this here because it is relevant to how a child or animal might think about mental states and processes as causes, both within itself and in others, though this document is more concerned with physical causation. Elsewhere I have suggested that biological evolution solved the philosopher's 'Other Mind's' problems
in its own way --- by providing genetically determined mechanisms and forms of representation for use in understanding the behaviour of other animals in terms of things like perception, knowledge, desires, mental capabilities.
How can I know that anyone else has a mind? How can I conceive of others having mental states and processes when all I observe is physical behaviour?
Whatever mechanisms are available for representing and reasoning about one's own mental states and processes, and causes therein, will also in principle, be usable for reasoning about others. (Some of the problems and difficulties to be overcome by such mechanisms are mentioned here in the section on hypothetical information and gaps therein.)
The details of such mechanisms need not all exist at birth, and it is very likely that they grow and develop during their use, but such growth requires some genetically determined starting point, which may be present in different forms with different potential in humans and other animals.
Initially human thoughts about mental processes in intelligent machines were something like metaphorical extensions of our ability to think about humans and other animals. But gradually, during the last 50 years or so we have acquired a new more solid understanding of virtual machines and causation therein, which is now used every day in designing, testing, debugging, explaining, and reasoning about ever more complex software systems many aspects of which, though not all, are deterministic.
Discussions with colleagues in the CoSy project and elsewhere have raised some questions about different sorts of requirements for understanding and manipulating objects in the environment. In particular, there are different architectural requirements for (a) physical control of movement and (b) understanding structures, processes and the relations between them.
(2) plan actions in advance of performing them, including considering different ways of achieving something and choosing a good one,
(3) explain how to do something: e.g. show the angle at which something must be held to fit into a slot, describe or show how big a gap needs to be for a particular thing to go through it, describe or show what shape something needs to be in order to perform some function (e.g. levering up the lid of a coffee tin),
(4) explain how something came about, e.g. why something broke, how something was caused to move, why a collision did not happen, etc.
(5) think about things that are not currently seen, e.g. where something must now be if it was moving at a constant speed and has gone out of sight behind something bigger, where someone foot is when only the upper half of the person is visible, say whether two people seen standing on the far side of a wall could be holding hands or not, etc.
In recent years there has been a lot of work on (a) (dynamic control) in robotics, e.g. in Rod Brooks' lab in the USA and other places. However it is very difficult to do and requires specialised robotic equipment with very sophisticated motors, sensors and sensory feedback, typically using force-control rather than position-control.
It is quite difficult to work on this without having mechanical engineering expertise and very expensive equipment.
However, it may be that that's not where we should be going if we are primarily interested in cognitive processes that are involved in the relatively rare, more sophisticated, animals that not only do things, but also can think or reason or communicate about what they do, i.e. have the reflective abilities of type (b).
For that we need to address different sorts of issues about forms of representation and reasoning which don't merely produce actions, but answer questions, make predictions, etc. They could be used in producing and controlling actions, but not necessarily with speed and fluency of the kind required for competence of type (a).
(Is this a distinction made clearly by psychologists and neuroscientists studying action? Personally I think that what Milner and others described as the difference between 'what' vs 'where' visual pathways was just confused, and was really concerned with the difference between vision used for tasks of type (a) and vision used for tasks of type (b).)
PDS:I think there is a wide-spread belief that is now current at least among some AI researchers, and I suspect many others, that PDS is true. Some even seem to go further and believe that the possession of the dynamic control capabilities is *sufficient* for all the sorts of tasks that we think need planning, prediction, explanation, etc. (I believe Rolf Pfeifer is an extreme example and possibly Tim van Gelder, in BBS 1998). I think there are echoes of this in the current interest in 'mirror neurons'.
It is not possible for a machine (or animal) to have reflective capabilities of type (b)
if it does not have an appropriate range of dynamic control capabilities of type (a).
PDS implies that in order to be able to reason, predict, explain, etc. features of movements of a certain type one must have previously been able to produce similar movements. I.e. deliberative and reflective capabilities depend on dynamic control capabilities.
One reason for thinking this might be a belief that processes of type (b) require mechanisms that *simulate* the processes that are being thought about, where simulation involves somehow running through the same control processes without performing the actions overtly. This might be a result of using feed-forward (predictive control) mechanisms to provide simulated sensory feedback when movements are simulated internally. Recent work by Murray Shanahan is an attempt to implement this sort of thing: http://casbah.ee.ic.ac.uk/~mpsha/ShanahanAISB05.pdf
Kinds of evidence for the falsity of PDS include the following facts:
If thinking about an action required the use of competence at achieving it then they could not think about doing anything, or understand an instruction or suggestion to do it, before being able to achieve it.
(Has anyone ever attempted a survey of animal species, including insects, fish, crustaceans, etc., asking the question: how many things do the young try to do and fail and later succeed in doing by developing new ways of combining sub-actions?)
One implication is that if we are to test cognitive mechanisms that do not depend on the low level fluent control of physical movements, then in addition to 'question answering' tests we will need physical tests that do not necessarily involve impressive motion. It could be that everything is slow and laboured, like human patients whose low level motor control mechanisms have been damaged but who are still able to use cognitive sub-systems to control their movements.
This also means that our requirements for robot arms can be much
simpler if we are not concerned with fast, fluent actions!
(Compare a normal human learning to control a mechanical digger by sitting in its cab and moving knobs and levers.)
Comments and criticisms welcome.