ann work flow
Jake
someone who understands them goes about working with a simple network.
I guess what I'd like to know is what is the work flow process is. How
does a user normally use a network? Is what I have below correct?
1) Choose a network
2) Give the network data
3) Train the network
4) Analyze the results? Using plots, graphs, etc...?
5) Run the network? This is equal to say, "production use" of the network?
Does that sound right?
Stephen Wolstenholme
wrote:
You should validate the results using data that the network has not
used for training. This can usually be done at the same time as
training.
Steve
--
Neural Planner Software Ltd www.NPSL1.com
Neural network applications, help and support.
Greg Heath
1. Analyze the data
a. rank and condition number of data matrices X,Y
and Z=[X;Y]; size(X) = [I N], size(Y) = [O N]
b. correlation coefficent matrix of Z
c. plot
1) yi vs xj and xi vs xj.
2) PCA plane projections
3) clusters (for classification)
d. edit redundant, irrelevant and incorrect data
e. normalize data and repeat a-d
(See my post on pretraining advice for newbies)
2. Obtain reference model performance
a. Obtain MSE00 for naive constant model out = mean(y)
b. obtain NMSE0 = MSE0/MSE00 and R0^2 = 1-NMSE0/MSE00
for linear model
3. Choose a network topology
4. Choose a training algorithm and initial parameters
5. Partition data trn/val/tst
To avoid overtraining, make sure the sizes of the
network and training set are compatible for the algorithm
(e.g, Early stopping, regularization or neither)
6. Design candidate networks
a. Train
b. Validate
c. If necessary, change network size and/or
algorithm parameters and go back to a
d. Test. Obtain NMSE=MSE/MSE00 and/or R^2 = 1-NMSE
e. Obtain multiple weight initialization designs
f. Either combine selected designs (ensemble) or
choose the best
7. For high confidence results repartition data and
repeat 5 and 6
8. Tabulate and plot results
Hope this helps.
Greg