Given a multilayer perceptron to be used to separate 2 classes. One has 2 design options.
1 -- Use one output node -- where one class is trained to give an output equal to zero, and the other class is trained to give an output of 1 on this node.
2-- Alternatively one could design a network having 2 output nodes; where one class gives an output of 1 on node #1 and zero on node #2, and the second class gives an output of 0 on node #2 and an output of 1 on node #1.
Which of these two options is considered better?
Also, I notice that the sigmoid function approaches zero but does not quite reach zero. Also, the same function approaches 1 but does not quite reach 1. Given this situation, how can one train a network to output a value of 1 or 0? Should one use something like 0.99 and 0.01 (as opposed to 0 and 1?)
