NS-FAQ-Running Networks

General
Algorithms
Platforms & OS Issues
Network Components
Network Construction
Editing Networks
File Management
Running Networks

Download the entire FAQ in Word 97(.doc) format.
If you don't have Word97....Download the Word Viewer.

Running Networks FAQ:

Q1. Why doesn’t anything happen when I hit the Start 
       button?
Q2. When using a test data set, why don’t I see anything 
       happening in the probes?
Q3. If I "fix" a component’s weights in the "Soma" property 
       page, are they permanently frozen?
Q4. How can I freeze a component’s weights while allowing 
       others to learn?
Q5.What’s the difference between randomize and jog in the
         control panel?

Q. Why doesn’t anything happen when I hit the Start button?
A. NeuroSolutions now keeps track of the total number of epochs trained. After a training run, if you wish to continue training, you must increase the epochs in the "Static" property page of the forward controller, click the "Zero Counters" button on the control toolbar, or reset the network and start over.

Q. When using a test data set, why don’t I see anything happening in the probes?
A. All probes have an access property page, where you can set the "Access Data Set", through a pull-down menu, to "All", "Cross-Validation", "Training", or "Testing". The default setting is "Training", and thus you need to set it to either "All" or "Testing".

Q. If I "fix" a component’s weights in the "Soma" property page, are they permanently frozen?
A.No, fixing a component’s weights simply means that they can’t be randomized. The component can, however, still learn and improve its weight estimates.

Q. How can I freeze a component’s weights while allowing others to learn?
A. For supervised components, set the associated gradient descent component’s learning rate to zero. For unsupervised components, set the learning rate of the component itself to zero.

Q.What’s the difference between randomize and jog in the control panel?
A. Randomize alters the weights according to a uniform distribution with the given mean and range. Jog alters the weights around their current value, using the given range. Jog can be useful during training to nudge a network out of a local minima.