Smaller nets may perform better: special transfer functions.
In an earlier study we stated sufficient conditions on the transfer function f of a feedforward multilayered neural net such that the output on R(n) defines the net up to trivial changes and smaller nets can have better performance on finite sets. Here we prove that 1/(1+e(-x)), tanhx, (1-e(-x))/(1+e(-x)) and invtanx satisfy these conditions.
Journal Article↗