Formalising the Use of the Activation Function in Neural Inference