next up previous contents
Next: The Well-designed Logical Robot Up: Experimental Possibilities Previous: AI

Psychological Experiments

Elizabeth Spelke [Spe94] describes a number of experiments that she and others have done to discover and verify innate mental abilities. The basic technique uses the fact that a baby will look longer at something surprising than at something that seems familiar.

Here's one that was first done in 1973 [Bal73] and was repeated by Spelke in 1993. There are experimental babies and control babies and the experiment has two phases. In the first phase the control babies are shown nothing. The experimental babies see an object go behind a screen and shortly another object emerges on the other side of the screen. The timing is 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 to get bored with it and stop paying attention.

In the second phase of the experiment the screen is removed. There are two variants. In the first variant, the first object strikes the second and knocks it onward. In the second variant the first object stops short of the second, but the second object takes off as though it had been struck. The control babies look at both variants for the same amount of time. The experimental babies look 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 removed, 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, i.e. are well-designed in that respect. For details see [Spe94].

That was an actual experiment. Now consider some possible experiments.

Suppose we want to determine whether some abilities concerned with a specific fact about how the world is organized is innate. We compare a baby's ability to use this fact compared to its ability to learn a fact about an environment constructed differently from our world but logically no more complex.

Here are some possibilities. Since I am rather innocent of the psychological literature some of them may already have been tried.

three-dimensional objects
I'm skeptical that a person's notion of a physical object is fundamentally visual. Here's an informal experiment I actually did. The subjects attempted to draw a statuette in a paper bag. They could put their hands into the bag and feel it as much as they wanted to. The quality of a subject's drawing, except for surface colors, was similar to what that subject would have produced looking at the object except in one case. The object was a statuette of an owl, and the subject who misperceived it as an angel produced an inferior drawing.

It would be worthwhile to use this and analogous techniques to explore people's concepts of three-dimensional objects. I would think that it is possible to investigate how babies perceive objects they are only allowed to touch and then see. The baby could feel an object in a paper bag and then see either the same object or a different object. The hypothesis is that the subject would regard seeing the same object as less surprising than seeing a different object.

anticipating the future
To eat when hungry doesn't require having in mind anything like a sentence. However, to know that one will be hungry 4 hours from now may require it. Maybe this is where humans and apes part company. Can an ape that is not hungry perform a non-habitual action, e.g. putting a key by an empty food box, in anticipation of being hungry later?

unethical experiment
A Lockean baby would do as well in flatland as in our space. Imagine arranging that all a baby ever sees is a plan of a two-dimensional room and all his actions move around in the room. Maybe the experiment can be modified to be safe and still be informative.

continuity of motion
The Lockean baby is brought up in an environment in which motion is discrete. Imagine that the baby's world is a Macintosh screen. Objects move without passing through intermediate points. The baby moves an object by clicking on the initial and final locations. The experiment is to determine how well a baby will do in such a world. This one might be tried with an animal.

attention experiment
If a baby is built to expect objects to behave as solids, then it will be surprised when objects appear to interpenetrate. It might pay longer attention to such a scene.

Babies might or might not find Escher-type drawings surprising.

geographical representation
Consider a maze with a glass top. Does it help an animal find food if it can walk around on the top of the maze before entering it? The top could have small holes that the smell of the food could get through. One psychologist opined that dogs would be helped and rats would not.

The experiment would test whether the animal can represent a scene by something like an image.

goal regression in animals
An animal seeks a goal but discovers that a precondition must be achieved first and undertakes to do it. Then it discovers a precondition for the precondition, etc. Suppose the animal has been trained to achieve B as a precondition for achieving when A isn't already true. It has been trained to achieve C as a precondition to achieving B when B isn't already true, etc. We ask how far the animal can carry the regression. Say the animals are dog which vary in intelligence, or at least vary in the ability to learn the tasks that humans teach dogs. We ask is there an innate limit for dogs or can smart dogs carry it farther than dumb dogs.

Susan McCarthy informs me that when a performing animal is taught a new trick, the trainer starts with the bow at the end and works backwards. I don't know if this is related to goal regression.

grammar via meaning
Many of the discussions of a child learning its native language seem to assume that the child learns grammar solely by observing grammatical regularities in speech and having its grammar corrected. Consider a child raised by an English speaking nanny whose native language is Spanish and is addicted to Mexican soap operas. It seems to me that this happens often enough so that observations could be made. The child would then hear a lot of idiomatic Spanish. It would be interesting to observe whether the child would be able to tell grammatical from ungrammatical Spanish sentence.

My conjecture is that grammar is learned as an auxiliary to meaning and is not separately represented in the brain.

next up previous contents
Next: The Well-designed Logical Robot Up: Experimental Possibilities Previous: AI   Contents
John McCarthy