I wrote an article titled “Neural Network Binary Classification” in the August 2015 issue of Visual Studio Magazine. See https://visualstudiomagazine.com/articles/2015/08/01/neural-network-binary-classification.aspx.
A neural network classifier makes a prediction based on predictor variables. For example, you could predict the political leaning (conservative, moderate, or liberal) of a person based on their age, education level, occupation, and so on. When the variable to predict can be one of only two values, the problem is a special case called binary classification. For example, the problem of predicting the sex (male or female) of a person based on their income, political leaning, and so on is a binary classification problem.
In my article I explain that there are two ways to implement a binary neural network classifier. One approach uses a single output node that always has a value between 0.0 and 1.0. If male is 0 and female is 1, then if the computed output node of the NN is less than 0.5 the prediction is male; if greater than 0.5 the prediction is female.
The other approach for binary neural network classification is to have two output nodes that sum to 1.0. If male is (0,1) and female is (1,0), then suppose the computed output node values are (0.652, 0.348). The computed output values are closer to (1,0) than to (0,1) so the prediction is male.
In my article I discuss the pros and cons of the two implementation approaches. I explain why I prefer the “two-node” technique but why most people prefer the “one-node” technique.