Download the entire FAQ in Word 97(.doc)
format.
If you don't have Word97....Download the Word Viewer.
Algorithms FAQ:
Q. How
does NeuroSolutions implement neural networks?
A.
NeuroSolutions adheres to the so-called local additive
model. Under this model, each component can activate and
learn using only its own weights and activations, and the
activations of its neighbors. This lends itself very well
to object-orientated modeling, since each component can
be a separate object that sends and receives messages.
This in turn allows for a graphical
user interface (GUI) with icon
based construction of networks.
Q. What
algorithm does NeuroSolutions use to train recurrent
networks?
A. NeuroSolutions
uses the back-propagation through time (BPTT) algorithm,
which "unfolds" a dynamic net at each time step
into an equivalent feed-forward net.
Q. How
does NeuroSolutions implement Radial Basis Function (RBF)
networks?
A. The centers and widths of the
Gaussian axons are determined from the cluster centers of
the data, which are found through an unsupervised
clustering algorithm. The weights from the Gaussian axons
to the output layer are then determined through
supervised learning with a desired signal.
Q. Can
I implement my own algorithms?
A.Yes, the easiest way to modify
NeuroSolutions is through Dynamic
Link Libraries (DLL’s),
available with the Developer’s Lite
and Developer’s
levels. Every component has default code, which can be
generated and edited from the "Engine" property
page, and then compiled with MS Visual C++.
|
