Messages

1

Off Topic / Re: Mechanical Neural Network

« on: December 18, 2010, 07:53:26 AM »

I hope I can help here. The most common type of neural network is feedforward, which means that input goes through the network and always produces an output. There is no looping and no halting problem; it is not turing-complete, but it is a universal approximator in the sense that it can make any mapping between a finite number of inputs and outputs. 0 can become 1 because there is always a dummy node in the network that outputs 1, and it can do XOR easily. Recurrent neural networks feed their output back into their input, but I think they're still not turing-complete unless their nodes' outputs are analog, as opposed to the more common binary threshold outputs, which would make them finite state machines without an infinite number of nodes.

A finite series of NAND gates is also a universal approximator.