Canonical PWL network and multilayer perceptron-like networks: A unified view

The authors consider the behavior of the most popular type of mapping networks, the multilayer perceptron-like (MLPL) networks in implementing or approximating functions, in terms of the canonical piecewise-linear (PWL) functions. They show that a MLPL network may be understood as performing a canonical PWL function or a PWL function which is a composition of the canonical PWL functions. The discussion further suggests a generalized class of the canonical-PWL (CPWL) networks, i.e., networks which perform a canonical PWL function or a composition of the canonical PWL functions, which includes all layered feedforward networks where the nonlinearity of the units is represented or approximately represented by a PWL function