Novelty of Neural Net Approach

[Bruce Krulwich]

I think that Teru Homma's recent message contrasting connectionist and
conventional computer systems shows the importance of distinguishing between
the hardware characteristics of a system and characteristics of its
reasoning.  The lack of such a distinction has resulted in several recent
articles in the media claiming that connectionism has solved all the
problems of AI, which I don't think is quite true yet.  Examples of such a
distinction follow:

>   Conventional Systems                    Neural Net (Connectionist) Systems

>8. have complex structure                  have rather uniform structure

This is true only at the elemental level.  Any particular connectionist
model may have complex (nonuniform) structure.  In fact, it seems to be the
case that connectionist AI models will have to scale up the complexity of
their representations before they will be able to handle any high level
cognitive behavior like goal-oriented activity, expectations, plans, etc.

>10. designed or programmed with            learn from examples;
>   rules specifying the system             may self-organize
>   behavior

Conventional computers can learn from examples just as well as connectionist
systems.  Conventional algorithms for inductive learning, which is basically
what back propogation, recirculation, Boltzmann learning (tentatively), and
competative learning all do, have been under development for over 20 years.
What connectionist models do provide is a uniform process model in which
inductive learning becomes easier.  Most connectionist learning algorithms
so far (most notably backprop) suffer many of the limitations that inductive
learning does, among them that on the one hand how well the system can learn
depends heavily on the representation chosen, but on the other hand if the
system is given a complex representation then it hasn't truly learned the
domain but simply learned to use the domain knowledge it started out with.

In summary I think that it's important to distinguish between the advances
in neural network systems and the advances in connectionist AI.  The two of
them have to discussed along different metrics and with different past and
future goals.


Bruce Krulwich

-------