Maybe I don't understand your question.
Is it that you simply have never heard of Chi-square before? If so,
get a copy of just about any elementary statistics texts (best bet, if
poor in mathematics: introductory text of statistics FOR PSYCHOLOGISTS,
which is likely to be much easier to read than other stat texts...).
It sounds kind of impressive, "perform validation tests on neural
networks", unless you are aware that the same old Chi-Square can be
used to "perform validation tests on cheese tasting", any other kind of
a test when you are trying to determine whether expected frequencies of
two or more categorized events differ from chance expectations.
For example, if we expect people to be indifferent to which cheese they
eat, we might expect 25 people of a group of 75 to eat cheese A,
another 25 to eat cheese B, and the remaining 25 to eat cheese C.
If we find, however, that only 15 ate cheese A, but 38 ate cheese B,
and 22 ate cheese C, can we still say that the cheeses are equally
preferred? Do a Chi-square test!
So, is it something like this? Neural networks should be able to sort
some kind of event or some kind of object into one pile, and the other
into another pile. We know that there are actually 220 of the one thing
and 190 of the other. If the neural network decides there are 200 of
the one and 210 of the other, is this close enough? i.e. not too far
from the expected frequencies, according to Chi-square? If so, the
neural network is validated??
Seems unlikely to me that it is really that simple a use that you are
reading about, but who knows?? could be!
Note that I do not say (for example) 40 percent of one and 60% of
another; use frequencies, not proportions or percents, else your
CHI-SQUARE TEST will be invalid...
(Of course, if you read further in the text, you may see a few other
statistical tests which might be more useful...))
F. Frank LeFever, Ph.D.
New York Neuropsychology Group
In <3793BC3A.839C01C7 at lex.si.ualg.pt> Nelson Martins
<a6242 at lex.si.ualg.pt> writes:
>>I'd like to pose you a question:
>could you help me about where could I find documentation about
>chi-squared tests to perform validation tests on neural networks?
>I've read pp. 67-74 from "Neurofuzzy Adaptive Modellinh and Control",
>from Brown and Harris, but I don't understand the meaning of the
>quantities
>\omega and L, in Eq. (2.25).
>>Thank you for your kind attention,
>>>>--
>Nelson E. Martins
>UCEH - Universidade do Algarve
>Lab. 2.70 Tel. +351 89 800949
>8000 Faro Portugal
