We know inputs (X), we know outputs (y). What we do not know is weights (w) that map the inputs to the outputs. We would like to find a configuration for weights such that the Xw will approximate to y as much as possible.