Neural Networks for Supervised Learning

Abstract

Quite commonly, we are faced with the problem of taking a vector x = (x1, … , xn) of inputs and producing a vector y = (y1, … , ym) of outputs. For example, in a classification problem, the x1, … , xn may be characteristics of an item to be classified, and the corresponding output could be a single y, the class label for that item. Hence, the task is to uncover a function g such that y = g(x). Of course, the mapping g may be nonlinear. Generally, we are satisfied if we can approximate the ‘true’ function g sufficiently accurately by a function f of some particular form, e.g., polynomial in several variables, where f has coefficients or parameters whose values we need to determine.