Note on Generalisation in ANNs
This note is relevant to both the Assessed
Coursework and to the (related) Programming exercise for
seminars in week 4.
In both these cases, the training set should include every possible pattern of 0s and 1s. Because these Boolean input
patterns are all so different from each other, with no common pattern,
you should not expect
generalisation; ie if you train on a subset and hope for generalisation
to unseen test cases, you will probably be disappointed.
This is in contrast to the discussion in Lectures, eg Lec 7, where we
considered case where generalisation is
possible. Eg with a training set of pictures of dogs, one
can reasonably expect a common theme to exist, so that an ANN, properly
trained, could generalise to an unseen test picture. But with the
Boolean inputs of the Assessed Coursework (and the seminar exercise),
this is not the case - these are generalisation-unfriendly types of
problems.