| Home |

General Intelligence :

Some examples

  1. Monotonous
  2. Fighting
  3. Ben throws the red ball to Izabel
  4. Kellie is lying
  5. more later...

Vision: pineapple

Vision: recognize pineapples. Illustrates the model-based and appearance-based approaches to vision.


Monotonous

Recognize if something is "monotonous". For example, a wallpaper pattern, a story, a job.


Fighting

Try to give the most complete and general definition of "fighting".

A special case is as follows:

John hits Mary;
Mary hits John;
John hits Mary;
[ repeat a few more times... ]
to infer: John and Mary are fighting.


Ben throws the red ball to Izabel

Ben throws the red ball to Izabel.

Question: After the event, is Ben still holding the ball?

This example is partly related to the concept of object permanance in cognitive psychology.

It may also involve the trick of "inverse pattern recognition". Normally, pattern recognition proceeds from the bottom up (sensory experience). For example, "Ben's arm moving outward and the red ball flying off Ben's hand" is recognized as "Ben throwing the red ball". What we need here is something in the reverse direction: we are given that Ben "throws" the ball, and we want to retrieve the fact that the ball flies away from Ben's hand, as part of the pattern of "throwing".

With a logic-based representation of patterns, doing the reverse of pattern recognition is known as abductive reasoning.


Kellie is lying

I ask Kellie out for a date, but she says she is busy. Then, I find her at the bar.

To infer: Kellie is lying.

At the beginning there should be the following goals in Working Memory:

  1. Go out with Kellie
  2. Make Kellie like me
  3. Have sex with Kellie
  4. etc...

The following sequence of inference occurs:

  1. At first I accept that being busy is a legitimate reason for not going out with me.
  2. Finding her at the bar means that she is not busy.
  3. Therefore, she is lying.

Details of the thought process

A.

Kellie says she is busyshe is busy unless she is lying
where "unless" is an operator that extends first order logic.

This involves 3 things:

  1. The use of nonmonotonic reasoning to handle the assumption that she is not lying.
    There are several ways of doing this, eg: circumscription, default logic, and the truth maintenance system (TMS) approach. See nonmonotonic reasoning.
  2. A rule allowing us to assume that what a person says is true given that that person is not lying:
    Say(person,p) ^ ¬Lying(person) → p
    How can this rule be learned?
  3. Kellie says "I'm busy". Here "busy" is taken to mean "busy with obligations / work" by convention, ie, Kellie is not busy watching TV etc. This context-dependent interpretation of "busy" should be provided by the natural language processor.

B.

To infer that: Kellie is at the bar → Kellie is not busy.

Here we need a special piece of knowledge stating that "going to a bar" is a kind of "free activity":

going_to_bar ⊂ free_activities.

This is a special operation that involves pattern recognition, which is not the same as recalling simple facts from memory. The difference is that a pattern may have an infinite number of referents, whereas memorized facts are finite. For example, one can expect to have a memorized fact like "going to a bar is a free activity" but there cannot be memorized facts for more unusual activities like "throwing boomerangs" or "juggling pencil sharpeners". To recognize an infinite number of possible "free activities", the Pattern Recognizer uses both instance-based and rule-based matching (see pattern recognition).

Finally we have another piece of semantic/generic knowledge stating that:

Kellie is busy → Kellie cannot be doing free activities

which allows to draw the contradiction.

Note: in this example we have ignored time and tenses, and the probabilistic values of statements.

What can be learned

  1. Pattern recognition is required in addition to inference.
  2. We need to be able to keep track of multiple assumptions (even contradicting ones, eg Kellie is lying / not lying) with different probabilities. Some truth maintenance systems can deal with this.
  3. One key aspect is the ability to recall relevant facts (or rules) from a huge knowledgebase.
  4. Many inference steps should be performed spontaneously (ie not goal-directed), with forward-chaining, even if they ultimately do not lead to useful conclusions.




Reference

| Home |

Mar/2006 (C) GIRG