Re: Bad results training NN

"MallRat" <[EMAIL PROTECTED]> wrote:
"I am training a BPFFN classier with a dataset of 240 vectors and
testing it with another 60. Input space dimension is 12 and output
space 5.

After some training I got 100% sucess on the training set but with the
testing set I got an awful 20%... What is wrong?

Shoud I use more training data?"


Others have suggested important considerations (especially the
possibility of overfitting), but one I'd add is the possibility that
the train/test split is not being performed properly.  Unless one has
a good reason to do otherwise (time series or stratification, for
instance), this split should random.  Keep in mind that many
"arbitrary" splits (such as the first 240 records vs. the last 60
records) will  not  be random in real data sets.

good luck,
Will Dwinnell
http://will.dwinnell.com