Analogous Reasoning Is Not Strong Enough

Marvin Minksy said in another discussion group:

"Generally, when one speaks of a formal
'mathematical' model, one is referring to a very
small handful of axioms (and assuming some rules
of inference). Thus for the theory of groups,
one has just 4 or five of these. Then one can
spend many lifetimes trying to find proofs for
statements about that model.

The brain has a structure that is based on the
interactions of many thousands of genes, whose
expressions are controlled by many other
thousands of genes. If one had descriptions of
all of these, it seems unlikely that one could
deduce anything useful from this?that is, by
making formal deductions in one person's head.
Instead, one would need to make many
simplifications, and then do extensive
simulations.

In particular, I would maintain that the most
common formal systems (namely logic and
probability) are simply too far from being
adequate to represent even the most common types
of commonsense thinking which, in my opinion, is
proposing solutions to problems by constructing
analogies to problems that one already can (more
or less) solve. I claim that formal logic is
not good at expressing such things, and
probabilities are simply too opaque; what one
needs are large databases of cross-connected
representations?such as the types that AI
researchers call "Semantic Networks" etc. Then,
one needs to apply to these, not 'mathematical
reasoning" but "reasoning by analogy," because
one needs to employ all sorts of heuristics to
guide the course of one's thinking.

In principle, mathematics is adequate for such
things. However, in practice, formal reasoning
requires so many simplifications that one cannot
trust the conclusions one draws from formal
models. This is why, at least in my view,
mathematical psychology has not made much
progress."


I agree that something similar to a Semantic Net is necessary for
intelligence, but I feel that Minsky is wrong about analogous
reasoning. Analogous reasoning is also inadequate, at least it is at
this time. Minsky pointed out that logic is inadequate because formal
reasoning cannot express the kinds of relationships between knowledge
that is required for intelligence, and I think that he was suggesting
that probability methods are not in themselves capable of revealing the
complex relations that are needed to produce explanations for observed
events. Although I think his reasoning about logic and probability is
strong, I do not believe that analogous reasoning can produce the
effects that we are looking for before the programmer has first
discovered how those effects may be achieved. If we discover the
insights that we need to know, and if those methods are found to be
feasible in computer programming, then we could probably use any number
of methods to teach the AI program. We could use analogous reasoning,
most of the reinforcement methods like the yes-no 20 question game (the
20000 question game as I call it), hybrid Neural Networks, Bayesian
Reasoning, numerical reasoning or a variety of other methods to produce
stronger intelligence. The point is that the problem has turned out to
be much more complicated than most experts had thought it was going to
be and until we find those missing theories and find ways to use them
in feasible programming constructs, nothing is going to work. I have
to qualify that statement. It is extremely unlikely that any of the
contemporary AI paradigms are going to produce stronger AI without some
additional theories about how knowledge, learning, and interactive
ideation can be represented in a computer program.

During the past years, and even during the past few months I have
gained some further insight about ways that a GOFAI model might be
turned into a more effective paradigm capable of producing greater
intelligence. However, I have not tested my theories, and right now I
am bogged down in the preliminary programming, so I cannot say whether
my theories are feasible or not. They may not be.

However, as I thought about Minsky's comments I realized that
analogous reasoning could implement some of the more fundamental
effects that I have been thinking about during the past few months.
But since analogous reasoning is subject to conceptual-relative
influence it is not in a form that would allow it to be used as an
quasi-elemental process capable of eventually producing independent
intelligence. Analogous Reasoning is a complex tool that requires
higher judgement as a prerequisite for its use. This makes it more
like a reinforcement tool, and even though it can be used to add more
structure than the typical reinforcement technique can, it is still
completely teacher-dependent. Without further insight about the AI
problem, analogous reasoning will not be an effective tool capable of
producing higher intelligence. Again, I have to qualify my statement.
If you happened to hit on the right kinds of heuristic relations that
were needed you might be able to use analogous reasoning to induct that
kind of knowledge into a strong AI paradigm, but that is not likely to
occur before someone methodically shows how it can be done.

I am not saying that Quine was wrong about metaphoric learning, but he
just did not have the advantage of the knowledge gained during the past
50 years to refine his viewpoint to a higher level.

Jim Bromer

.

Relevant Pages

  • Re: Analogous Reasoning Is Not Strong Enough
    ... mathematics is adequate for such ... Analogous reasoning is also inadequate, ... that probability methods are not in themselves capable of revealing the ... in feasible programming constructs, ...
    (comp.ai.philosophy)
  • Re: Analogous Reasoning Is Not Strong Enough
    ... reasoning" but "reasoning by analogy," because ... in mathematics working ... Analogous reasoning is also inadequate, ... in feasible programming constructs, ...
    (comp.ai.philosophy)