Neural Network

[censix]

Well I would say that, whatever Nick, yourself and others have in mind for implementing 1-layer MLP training on
the Epiphany architecture, if that works, it would already be HUGE success! And it would be very useful indeed.

[roberto]

Thinking deeper, there is alternative to Back Propagation algoritm to train a Neural Network: to use Genetic Algoritm to evolve a population of weights till reach the element that sutusfy the NN. This is a interesting path to explore, cos it is demostrate that can give great speed up to train and in this case, parallela is a overkill because AG need a simple classic cpu to be runned. I played with GA in past and they can give extraordinary results (few minutes of calculation) where a classic algoritm "brute force" need centuries to produce a result. In this case, the Back Propagation algoritm is concettually a brute force system, no matter if multy-layer-few-neurons-per-layer or single-layer-with-many-neurons NN.

[roberto]

[over9000]

[roberto]

[over9000]

[censix]

[CIB]

[over9000]

[nickoppen]

Hi Guys,

Here is the second part of the design - Training. It is a little more complicated that the feed forward pass so please give it a read and let me know if I've missed anything.



Thanks,

nick