Scaled/unscaled data in SNNS question
(Që… ð@ÄØð@ÄÔ)
Thu, 06 Feb 1997 13:45:44 -0500
I have been training a neural network with 1 input node, 3 hidden nodes in one
layer, and 1 output node. My data consists of inputs ranging from 0-15 and
outputs ranging from 0-15000. So far, I have been scaling the data by dividing
the outputs by a constant such that they all fall within the 0-1 range. I
would like to be able to avoid this procedure so that I can use the raw
(unscaled) inputs. I thought that I could accomplish this by simply adding
another layer to the end of the net. This new net would have 1 input node, 3
hidden nodes in the first hidden layer, 1 hidden node in a second hidden layer,
and 1 output node. I used a logistic activation function for the hidden units
and changed the output activation function to be act_identity. I thought that
the additional weight and identity node on the end of the network would serve
as an auto-scaling factor, but was quite mistaken. Instead, the new network
performs horribly. It converges on producing outputs of either 0 or a large
positive number which turns out to be the weight of the connection between the
last hidden node and the output node. I would appreciate any help that the
group can offer. Thanks in advance.
-Jeff Thieme
thiemero@pilot.msu.edu