> From: "Marshall, Helen" <hvm195@soton.ac.uk>
> Date: Fri, 24 May 1996 10:32:59 GMT
>
> Neural nets can be constructed physically but are often simulations
> because what is important is that they do what we do, not do it in the
> same way we do it.
Two issues conflated here: (1) nets need not be like real neurons;
(2) nets can be simulated by symbol systems. But, yes, it is important,
eventually, that our models do it the same way we do it; for now,
though, the models can do so little, it makes more sense to scale up
their capacities to a closer approximation of our own than to worry
about whether they are brainlike. In the end, doing it all (the Turing
Test) may be enough of a constraint to rule out anything but the way our
brains actually do it. If not, there's probably no way to know whether
the difference matters.
> However, for nets to give a realistic impression of
> what we do they should have similarities with the brain because our
> cognitions occur in our brains.
Sounds like you are now disagreeing with what you said earlier.
> Nets are like the brain in several
> ways, for example, both have interconnected units - in the brain these
> are neurons which send messages across a synapse and in nets these are
> connecting nodes; both have positive and negative connections which
> means that the neurons/nodes can have either excitatory or inhibitory
> effects on surrounding ones.
Most nets only have one kind, though.
> Both nodes and neurons pass activity to
> each other. The brain does this through action potentials and nets do
> it via a transfer function which means spreading activation, i.e. when
> one node is activated it brings those it is connected to into a similar
> state of activation. This process is called propogation. Also, like the
> brain, the connections in nets can become weaker or stronger and can be
> coded locally or in a distributed way. A final similarity between nets
> and the brain is they both have the ability to learn new things.
>
> However, there is evidence of differences between the brain and nets.
> For example, nets are arranged in layers but the brain is not
We don't know whether there are "layers" of "hidden unit" in the brain,
in the neural-net sense; but there are certainly plenty of
interneurones, neither sensory nor motor, i.e., neither "input-layer"
nor "output-layer."
> and they
> are very different in terms of biology.
What does that mean (to kid-sib)?
> Also, nets are not efficient at
> logical reasoning or language whereas the brain can cope with these
> well. Nets are similar to the brain in some respects but not others so
> perhaps as we learn more we can develop them so that they can do more
> of the things we do.
Not bad. For a higher mark, sort out the conflated issues, and relate to
the bigger questions, about computation, cognition, and categorisation,
as well as reverse engineering.
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:44 GMT