$($X$1,$X$2)$ are input features and target classes are either $+1$ or $-1$ as shown in the figure. x$2$ is never negative. A fully connected multilayer perceptron network is used for classification with $100\%$ accuracy. Step activation functions are used at all nodes, ie$,$ output$=+1$ if total input $>=$ bias b at a node, else output$=-1$