Time invariant feed forward
File names:
NumberDetection.mbn and NumberDetection.mbl
The example shows a time invariant Feed Forward net that is intended to detect
numbers between 0 and 9 in a matrix of input neurons.
There is also a lesson included in the examples that has the same name as
the net. The net is already trained to reproduce this lesson, still it is not
really good at interpreting new input patterns. You might want to play around
with that net using MemBrain's "Paint Brush Selection" feature and also add
more patterns to the Lesson by doing so. If you don't know about the 'Paint
Brush Selection' method available in MemBrain then search the MemBrain help file
for the corresponding entry.
Also you could modify the net architecture and see if you can get better generalization
results.
Note: This is a quite big net and performance is significantly improved if
you deselect the options <View><Show Links> and <View><Show Activation
Spikes on Links>.
The net has been trained using the 'Std BP with Momentum' teacher. However, since
this example has been created MemBrain has been equipped with more advanced
learning algorithms, e.g. the 'RPROP' teacher which can be used for this net without
changing any of the teacher's default parameters.
However, if you still want to use the 'Std BP with Momentum' teacher you can use the
following settings for this teacher:
If you click on the button 'Advanced':
Note that training will take significant time because of the size of the Lesson and the large number
of links that have to be trained!
Be sure to adjust the setting <Teach><Set Teach Speed> to a value of '0' as this will result in the
fastest possible teaching speed.
Copyright © 2003 - 2007, Thomas Jetter