We describe a method for learning a piecewise affine approximation to a mapping f : ℝ<sup>d</sup> → R<sup>p</sup> given a labeled training set of examples {x<sub>1</sub>,..., x<sub>n</sub>} = X ⊂ ℝ<sup>d</sup> and targets {y<sub>1</sub> = f(x<sub>1</sub>), ..., y<sub>n</sub> = f(x<sub>n</sub>)} = Y ⊂ ℝ<sup>p</sup>. The method first trains a binary… (More)