Prev: new image
Next: Edit a plot in a GUI
From: Nina on 8 Jul 2010 10:27
i can get output weights, but
i want to interprete weights in the hidden layer?
regards
nina

"Carlos Lopez" <clv2clv(a)removeThisadinet.com.uy> wrote in message <eef61e4.0(a)webx.raydaftYaTP>...
> You pose two different problems:
> a) > I tried to do this, but MatLAB can't work with 200 neurons
> (variable size too large error).
> I have never had such a "large" problem. From my viewpoint, 200 is
> not too large a problem, so please report version, operating system,
> etc. to figure out what is happening. Are you shure that the problem
> is within the ANN toolbox?
>
> > If I use less than 200 neurons, is there a
> > way to interpret the weights in the manner I described above?
> Interpretation of the weights is somewhat tricky. Indeed, I think
> that it is an open research area. You only briefly describe your
> problem, but I could say that there should be some hidden layer,
> which needs not to have 200 neurons. The weights could be interpreted
> not by the "input" weights, but through the output ones. Thus, if you
> have just two neurons in the hidden layer (and assuming that you are
> producing a single value as output; no information about this is
> provided) you could "interpret" the ANN analyzing the weights of the
> output layer.
> In other words: if the final output is composed of
> b3+w3*f(w1*x1+w2*x2), you will find that w1>>w2 (so the neuron
> #1 is significantly more important than the #2), etc.
> I have worked in this topic in the past; let me know if you want some
> references.
> Regards
> Carlos
 | 
Pages: 1
Prev: new image
Next: Edit a plot in a GUI