All the classification and regression procedures we have seen till now assume that a reasonable set of features is available. If the procedure didn’t work well, we needed to use domain knowledge, problem insight, or sheer luck to obtain more features. A neural network offers an alternative option: learn to make good features from the original signal. A neural network is made up of units. Each accepts a set of inputs and a set of parameters, and produces a number which is a non-linear function of the inputs and the parameters. It is straightforward to produce a k way classifier out of k units.