It's primitive; it's dumb; it's brittle--but it's AI.
eric at GRUVER.NET
Mon Jul 5 18:20:05 EST 1999
Sergio Navega wrote:
> Today I see no point in doing anything related to AI without a
> strong biological plausibility.
This strays a bit from the discussion, but let me point out
that AI (Artificial Intelligence) and ANN (Artificial Neural
Networks) were pretty much separated into two distinct
fields by Marvin Minsky many years ago. From my point
of view, there is little in AI today which has much
biological plausibility. In ANN, there is both
biological and non-biological approaches. In this
newsgroup, the biological approach is more on-topic.
In the non-biological approach, a neural network,
while originally a model for neural computation, is a
mathematical tool that is useful for making predictions
based on patterns of input. For these types of problems,
the original biological basis is of little importance
compared to the results gained. For example, using a
neural net to analyse weather patterns, predict
recurrence of breast cancer, or steer a driverless
vehicle, the biological basis of the neural networks is
really not important. In many cases, there are other
approaches without any biological basis that may perform
as well or better than the neural network approach.
A more biological approach using neural networks is to use
the neural networks to model the organization of the brain.
> One line of argument is that we must
> follow the path of the natural intelligences until we grasp what
> are the core points of intelligence, because we *still* don't know
> what they are. Only after that we will be able to "propose" new
> methods and algorithms to enhance biological intelligence with
> functionally equivalent (but better) processes.
> This does not preclude experimentation: often we'll have to create
> "strange" things, based on unlikely methods. What is important is
> not to lose the goal of obtaining plausible results, comparing
> our results with children's ways of learning and adult's methods
> of tackling new problems.
But it is probably impossible for AI to mimic biological
intelligence unless the architectures of the computer and
the brain are quite similar. If nothing else, the very
different architectures of the computer and the brain
demand very different approaches to intelligence. You
seem to be arguing that AI should work the same as biological intelligence in
spite of these architectural differences.
What is efficient for the brain is often very inefficient
for the computer and vice versa. For example, in a game
of chess, the human expert can quickly zero in on a limited
number of moves and analyze them to a pretty good depth.
In contrast, a computer will analyze a far greater number
of moves but to a lesser depth. The use of heuristic rules
to determine the more important moves can be used to
reduce the number of different moves to consider and allow
them to be analyzed to a greater depth, but the human
expert will still consider a small fraction of the moves
considered by any good computer program. In other words,
based on the architecture, different approaches to the
game are necessary.
> In summary, the road to intelligent systems is too elusive for us
> to waste time with implausible and risky methods: we ought to follow
> the steps of our brain, the best example of intelligence on Earth.
When we can build a computer based on biologically based neural
nets so that the architectures of the computer and the brain
are similar, then we might be able to have a more biologically
based AI. Until then, I don't really see how it would be
More information about the Neur-sci