THE WELL-DESIGNED CHILDJohn McCarthy
http://www-formal.stanford.edu/jmc/September 28, 2007
- • The world into which a human is born is complex. W
- discuss some of the complexities, concentrating on one—
- 3-d semi-permanent, movable objects.
- • Evolution has equipped a baby with some innate kno
- edge of the world. That works better than starting as
- blank slate.
- • Consider a well-designed logical robot child (WDC).
- Mostly it will have the innate abilities we conjecture real
- children have, but sometimes we can do better.
THE LOCKEAN BABY
- • John Locke 1693: The baby starts out as a blank slate.
- It builds its knowledge by inferring associations among
- • Since 1950 people proposed to start with a Lockean
- blank slate baby machine and have it learn from its exp
- • Starting from the blank slate, I’ll bet it’s a lengthy
- process to infer the existence of 3-d objects.
I don’t- think anyone has made an AI system that can do it.
- • A billion years of evolution has provided us with prett
- good prejudices about the world far better than a blank
- slate. Our robot child should also have them
THE WORLD IS COMPLICATED- • The world’s structure is not directly describable in terms
- of the input-output relations of a person. The basic struc-
- ture of the world involves elementary particles on time
- scales of 10−25 seconds, but intelligence only evolved
- structures of more than 1023 elementary particles.
- • Even at the level at which a small child can perceive
- the world is extremely complicated. Here are some of
- • Reality and appearance Animals and humans don’t p
- ceive the structure of our environment directly. Senses
- have evolved to give partial
information about objects- and their relations.
- • The world is 3-d, but our senses react to surfaces.
- • semi-permanent objects Much of the world consists
- three-dimensional objects that have masses, moments,
- compliances, hardnesses, chemical composition, shap
- outer surfaces with textures and colors, are often made
- identiﬁable parts which sometimes move relative to each
- other. A particular object can disappear from perception
- and reappear again.
- Note that the structure and location of an object in
- world are more persistent than its appearance and lo
- tion in the visual ﬁeld or relative to the hands.
- • temporal structure The environment of a child has
- complex temporal structure. Some items change in sec-
- onds, others last for hours, days, or years. Babies
- in the present. The concepts of tomorrow and yesterda
- are not learned right away.
- • causality Events cause changes in objects and their
- lations and cause other events.
STILL MORE COMPLICATIONS- • solidity Objects are solid and do not ordinarily penetrate
- one another. Some are rigid and some are ﬂexible.
- • gravity Unsupported objects fall to a lower surface.
- • kinds of objects Objects have kinds, and objects
- the same kind have properties associated with the kind.
- Babies are ready very early to learn what kinds there a
lemons, belong to natural kinds. The objects of- • natural kinds Many of the objects a child encounters,
- natural kind have yet undiscovered properties in common.
- Therefore, a natural kind is not deﬁnable by an if-and-
- only-if sentence formulated in terms of observables.
- • relations Objects not only have individual properties
- and belong to kinds, but objects and kinds have relations
- with one another. At least some ternary relations such
- as betweenness are basic. Also “A is to B as C is to
- seems to be basic.
In its numerical use, it reduces- the equality of two fractions, but the quaternary relation
- seems to be basic in common sense usage.
WHAT DO BABIES KNOW AND WHEN?- • Q: If the world is so complicated, how can babies
- anything purposeful?
- • A: They know simple cases of phenomena.
- • There’s good psychological evidence that they have
- innate knowledge of solid objects that continue existing
- even when out of sight. How do they remember an object
- that has gone out of sight so as to recognize it when
- • They are ready to ascribe purposeful action to humans
- and animals and to try to inﬂuence them.
- • the principle of mediocrity (from the astronomers) Our
- child is like other children. This lets it reason in b
- • persons Some objects are animate and have purposes
- analogous to those of the child. They can be inﬂuenced
- but are sometimes to be feared. Defects in innate under-
- standing of persons, e.g. autism, harm the child.
WHAT DO WE WANT IN A WELL-DESIGNEDROBOT CHILD?
- • Distinguish appearance from reality
- • natural kinds Natural kinds don’t have if-and-only-if def-
- initions. There may always be more properties to
- learned. To a small child, all kinds are natural. The
- robot child should think in terms of natural kinds.
- • three-dimensional objects These are more stable than
- the perception of them by any sense.
- • perceive motion as continuous
- • actions and their eﬀects
- • recognize parts Recognize parts of an object and their
- relations to the others. It would be interesting to have
- grammar of 3-d physical structure analogous to that
- • focussed curiosity
- • grammar of goal regression To do A, I need to do
- ﬁrst, and to do B, I need to do C ﬁrst.
- • introspection Children begin to do this by age 3, do
- well by age 5. The WDC needs it.
THE SPELKE EXPERIMENT- This psychological experiment exhibits abilities of human
- babies we want in the WDC.
- Elizabeth Spelke described a number of experiments that
- she and others did to discover and verify innate mental
- abilities. The technique uses the fact that a baby
- look longer at something surprising than at something
- that seems familiar.
- Here’s one that was ﬁrst done in 1973 and was repeated
- by Spelke in 1993 with two months old babies. There
- are experimental babies and control babies and the
- periment has two phases.
- babies are shown nothing. The experimental babies
In the ﬁrst phase the control- an object go behind a screen and shortly another object
- emerges on the other side of the screen. The timing
- such as would be appropriate if the ﬁrst object struck
- the second object and knocked it from behind the screen.
- The babies are shown the phenomenon enough times
- get bored with it and stop paying attention.
- In the second phase of the experiment the screen is
- moved. There are two variants. In the ﬁrst variant,
- ﬁrst object strikes the second and knocks it onward.
- the second variant the ﬁrst object stops short of the sec-
- ond, but the second object takes oﬀ as though it had
- been struck. The control babies look at both variants
- the same amount of time. The experimental babies lo
- longer at the second variant.
- The conclusion is that the experimental babies inferred
- that the ﬁrst object had struck the second when the event
- occurred behind the screen. When the screen was
- moved, they were not surprised when the expected event
- was shown to occur but were surprised and looked longer
- when this expectation was not met.
- The conclusion is that babies have innate expectations
- about dynamics. For details see Spelke’s 1994 article
- Cognition, Initial knowlege: six suggestions.
THE WELL-DESIGNED CHILD AS A LOGICALROBOT
- • Not even a sketch of a design—just some ideas.
- • Appearance and reality
- Appears(appearance, object) is too simple except in a lim-
- ited context, but children think in limited contexts.
- Holds(Appears(person, appearance, object), s) says more
- is suitable for referring to the child from the outside.
- • The correctness of a child’s beliefs and references
- objects depends on context. Children change context
- • When looking at a child’s ideas from the outside,
- can use the theory of contexts as objects introduced
- my Notes on formalizing context. In that theory Ist(c, p
- true if the proposition p holds in the context c. V alue(c,
- is the value of the expression exp, e.g. Color(Block1),
- context c.
- • V alue(Belief
s(Child1, S0), Color(Block1)) = Red asserts
- that in the context of Child1’s beliefs in situation s, Blo
- is red. The context theory lets us enter the context
s(Child1, S0). Then we have Color(Block1) = R
THREE CHALLENGES AND A CONCLUSION- • Make a robot baby that can do the Spelke experiment
- honestly. But what’s honestly? It may start knowing
- about collisions and knowing about occluded objects,
- at least it shouldn’t start with knowledge combining
- • Its contexts should change like those of a baby.
- • It should properly relate 2-d appearances to 3-d realit
- and also relate tactile appearances to reality. Here’s
- • What babies know presents challenges to AI—to AI
- all schools.
- • This talk is partly based on the manuscript