Locally connected neurons
First of all, encog is awesome. I used it last year on a small project and it made my live a lot easer!
While reading an a paper about face detection I was wondering about two things (... and I might have misunderstood the paper so I might be writting rubish here).
Is it possible to setup connections only between specific neurons from one layer to the next? I guess that would mean setting some weights to a fixed value of 0 which cannot be changed by training.
e.g.
i1 --\
i2 ---> h1 ...
i3 --/
i4 --\
i5 ---> h2 ...
i6 --/
Cascaded Face Detection Using Neural Network Ensembles:
http://www.hindawi.com/journals/asp/2008/736508.html
RSS
YouTube
Twitter
Facebook
I need the same thing...
I also need exactly what rednoah is describing, arbitrary connections (determined by me) between layers. Some neurons in layer 1 connected to some in layer's 2 and 3, etc. Would it be possible to accomplish this by making each neuron its own layer? I'm just starting w/ encog, so maybe I'm wrong. At any rate, this would be a cumbersome way to do it, if possible.
Encog looks awesome. I have been using SNNS for quite a while, and would love to migrate to Encog. Problem is, I need this functionality. I will check in periodically, and when this is implemented I'll make the switch.
Thanks for providing encog. It is quite a service you provide the community!
Sincerely,
John
This will be included in 2.4
In a way, it is already a part of 2.4 in that 2.4 supports NEAT, which sort of evolves "layerless" neural networks.
It may be possible to do this in Encog with single neuron layers, I have not tried. The most optimal way is to just represent it as a layer-based feedforward network, but "zero-out" the connections not needed, and have some way for Encog training to leave them alone. Which, as Sema stated, it currently will not. It would try to train a zero connection.
This has been requested enough times that I am going to make sure this is implemented in 2.4. All of the other 2.4 work is progressing nicely, so it should fit in just fine.
don't know of a way to do that in Encog
It could be done by setting the weights for those neurons to zero, but Encog WILL try to train them. But now that I think of it, there is going to be no gradient on a zero weight, so propagation training would very likely leave them alone. annealing and genetic would for sure mess with those weights.
Eventually, we need to specify a flag or something that tells Encog to not even attempt to train something with a zero weight.