Re:cPerceptron: A neural network (Simple Perceptron) impleme...
by illicium on Saturday December 15, 2012 @03:20AM
(#4924920)
>a perceptron cannot perform a XOR on inputs
Multi-layer perceptron networks, however, can.
Re:cPerceptron: A neural network (Simple Perceptron) impleme...
by jacques_chester on Saturday December 15, 2012 @03:24AM
(#4924928)
Right. According to my old notes you need:
* at least one layer to handle linearly separable functions,
* at least one hidden layer to handle continuous functions, and
* at least two hidden layers to solve continuous and discontinuous functions.
1. simple to code up
and
2. not very useful.[1]
You still need to pick the sort of neural network you're building (there are literally hundreds in the literature) before anything of consequence is going to happen.
The reason to use C/C++ libraries is that neural nets are slow. Slow slow slow. Computational intelligence / nature inspired computing is extremely cool, but the smart thing is to avoid it like the plague unless it is the only workable alternative.
[1] To the point that neural network research came to a screeching halt for some time when it was proved that a perceptron cannot perform a XOR on inputs.