Download the entire FAQ in Word 97(.doc)
format.
If you don't have Word97....Download the Word Viewer.
File Management FAQ:
Q. How
do I feed a test set through the network?
A. First, go to the
"File List" property page of the input file,
click "Add…", and choose your test set.
When the "Associate File" panel appears, choose
"Testing" under the "Data Sets" list.
If your test set has an associated desired file and you
want to view error information, repeat the above steps
for the desired file. Then, go to the "Static"
property page of the forward controller, and choose
"Testing" in the "Active Data Set"
pull down menu. This automatically turns off learning.
From the "Access" property page of the output
probe, switch the "Access Data Set" to
"Testing". You may then run the test set.
Q. How do I
split a file into a training set and a test set?
A. Add the file as usual
from the File List property page of the File Inspector.
It will be added as a Training data set by default. Click
"Customize…" and then, within the
"Data Segmentation" area of the resulting
pop-up window, click the"Segment" box, and
choose the offset and duration of the Training set. Then
add the same file to the Data Set list, but this time as
a test set (see the previous question). With the Testing
file highlighted in the Data Set list, click
"Customize…"again, and then the
"Segment" box, and set the offset and duration
to a different region of the file. These procedures must
be applied to both the input and desired files. In
addition, make sure you change your probe’s Access
Data Set (see Question2 on this
page).
Note that this technique can only
be used to split files into contiguous regions. To split
a file randomly, use NeuroSolutions
for Excel™ , or some other external pre-processor.
Q. How can I
train a network to do prediction of a data set without
using multiple files?
A.Add the input file as usual from the File List
property page of the File Inspector. Click
"Customize…" and then, within the
"Data Segmentation" area of the resulting
pop-up window, click the "Segment" box. Leave
the offset at zero. If you are predicting P steps in
advance, and your file has N exemplars, enter the
difference N-P as the duration. Then add the desired file
and click "Customize…". Click the
"Segment" box and enter P as the offset and the
difference N-P as the duration.
This process of setting up prediction is much simpler
in the NeuralWizard.
Consider using it if your network is one of the default
architectures it constructs.
Q. What is
the normalization file?
A. NeuroSolutions can scale
and shift your data so that it fits in any range. This
can be set from the "Stream" property page of
the File Inspector. The normalization file is where
NeuroSolutions stores the scale and bias parameters. You
can view, select, and resave the normalization file under
a different name from the "Data Sets" property
page.
Q.How do I
denormalize the network output?
A. Place a probe on the
network output, and go to the "Probe" property
page, and check the box "Denormalize from
Normalization File". Then choose the file that is
used to normalize the desired output.
Q. What is
the meaning of the various NeuroSolutions extensions?
A. The extension
"nsb" indicates a breadboard, "nsm"
is for a macro, "nsn" is used for a
normalization file, "nss" is a symbol file, and
"nsw" indicates a weight file.
|
