Custom Neural Network Structure
bibstha
according to http://vasc.ri.cmu.edu/NNFaceDetector/ (Rowley et al).
Although it is an old technique, i want to compare its performance
with modern techniques.
My problem is, they have described the neural net structure such that
not all of the inputs are connected to the hidden layers.
Say i have 400 input pixels, 80 initial pixels are connected to the
first hidden layer, another 80 to second hidden layer and similarly to
5th 80pixels connected to the 5th hidden layer. The hidden layers then
are connected to the output.
Can you suggest a neural network library that is flexible enough to
define the internal connections between hidden layers and input
layers? I have tried using FANN but it doesnot allow changing internal
connections.
Also can this type of Neural Network Structure be simulated in matlab?
I would code the neural network myself but wouldn't like to reinvent
the wheel if its already done.
Regards
Bibek Shrestha
Greg Heath
> Im trying to implement the a Neural Network Based Face Extractor
> Although it is an old technique, i want to compare its performance
> with modern techniques.
>
> My problem is, they have described the neural net structure such that
> not all of the inputs are connected to the hidden layers.
>
> Say i have 400 input pixels, 80 initial pixels are connected to the
> first hidden layer, another 80 to second hidden layer and similarly to
> 5th 80pixels connected to the 5th hidden layer. The hidden layers then
> are connected to the output.
>
> Can you suggest a neural network library that is flexible enough to
> define the internal connections between hidden layers and input
> layers? I have tried using FANN but it doesnot allow changing internal
> connections.
>
> Also can this type of Neural Network Structure be simulated in matlab?
>
> I would code the neural network myself but wouldn't like to reinvent
> the wheel if its already done.
Since none of the hidden layers are connected to each other, it is
equivalent
to a net with one hidden layer wirh specified input to hidden weights
fixed to zero.
Hope this helps.
Greg
bibstha
the weights to zero? it will eventually change.
On Nov 3, 1:13 pm, Greg Heath <he...@alumni.brown.edu> wrote:
> On Oct 29, 4:30 am, bibstha <bibekshres...@gmail.com> wrote:
>
>
>
> > Im trying to implement the a Neural Network Based Face Extractor
> > according tohttp://vasc.ri.cmu.edu/NNFaceDetector/(Rowleyet al).
GaryLynn
>define the internal connections between hidden layers and input
>layers?
NeuroSolutions includes an Arbitrary Synapse componponent that will
allow you define the individual connections between layers. You can
download an evaluation version from:
http://www.nd.com/neurosolutions/download.html
Gary Lynn
NeuroDimension, Inc.
tomhoo
> Although it is an old technique, i want to compare its performance
> with modern techniques.
>
> My problem is, they have described the neural net structure such that
> not all of the inputs are connected to the hidden layers.
>
> Say i have 400 input pixels, 80 initial pixels are connected to the
> first hidden layer, another 80 to second hidden layer and similarly to
> 5th 80pixels connected to the 5th hidden layer. The hidden layers then
> are connected to the output.
>
> Can you suggest a neural network library that is flexible enough to
> define the internal connections between hidden layers and input
> layers? I have tried using FANN but it doesnot allow changing internal
> connections.
>
> Also can this type of Neural Network Structure be simulated in matlab?
>
> I would code the neural network myself but wouldn't like to reinvent
> the wheel if its already done.
>
> Regards
> Bibek Shrestha
If your NNet package has the ability to prune (permanently eliminate
weights), you can use this technique to get fancy connections between
layers.
For your application, you would lay out partitions of the weight
matrix between the input and hidden layer so you only get the
connectivity you desire and then use pruning to zero the partitions
where you do not want connectivity/weights.