I am undertaking a course in Neural Networks and the Professor introduced us to the XOR problem. I understand the XOR problem is not linearly separable and why we need to employ Neural Network for this problem.
However, he mentioned XOR works better with a bipolar representation(-1, +1) which I do not understand. I have looked into resources in different websites yet I can't really understand the reason.
I am wondering why a bipolar representation would be better than a binary representation? What's the rationale for that argument?
The weight deltas of input nodes involve input values. When using the binary representation, an input node may have value 0, meaning that its weight delta is 0. In other words, this node can't 'learn' anything when this input vector is applied.
By contrast, if a bipolar representation is used, this can be avoided because the input nodes never have value 0. It means input nodes can always learn and thus helps the training converge faster.