> Date: Fri, 24 May 1996 12:09:07 +0100 (BST)
> From: "Lyne Katherine" <kml295@soton.ac.uk>
>
> Neural nets are suggestions for the woking of the brain.
For kid-sib: What are neural nets?
> It was
> suggested that, due to a combination of levels and systems, neural nets
> could perform tasks in the same way that a human brain does, including
> learning and producing output.
What did that mean?
> Pinker however argued that neural nets
> could not be an accurate represention of the brain as they were unable
> to cope with applying rules.
Applying rules or learning them?
> The nets are capable of recognising and
> working with examples that they have already met but they are unable to
> apply rules to new, unseen examples.
Pinker suggested that perceptrons could only memorise special cases;
they could nto learn regularities, rules.
> For example, humans are able to
> cope with the rules of grammar.
This example applies only to past-tense transformation rules, not to
grammar in general.
> In applying the rules, a human would
> say in a sentance that "he walks" present tense, but "he walked" past
> tense. They would also be able to remember that "I go" goes to "I went"
> with an irregular. Pinker argued that this ability to recognise and
> apply special cases was the only quality that hu8mans and neural nets
> share - because nets are unable to apply the normal rules of grammar or
> any other example.
Again, the issue is not one of applying but of learning from examples
and feedback about what's right and wrong.
> This could be due to lack of feedback - the net can
> never know that the rule it has applied is correct. It can only work
> with examples that it has seen before and recognises.
This seems to confuse the poverty of the stimulus, which concerns rules
of Universal Grammar, with past-tense formation, which is not part of
Universal Grammar and does not suffer from the poverty of the stimulus.
Pinker's critique is that neural nets can only memorise special cases;
they cannot learn rules. For that you need a symbol system. The critique
only applied to the perceptron: A net with hidden layers can learn
past-tense rules and other rules without difficulty.
This needs to be integrated with bigger questions about language and
symbol systems.
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:44 GMT