Cubic spline approximation of transfer functions for speeding neural networks performances

We propose the approximation of transfer function tanh (i.e. the Hyperbolic tangent) by cubic splines. Thus we save many multiplications and a division required for the standard double precision evaluation of this function. The cost we have to pay is to admit at most 2 decimal digits of inaccuracy in the final result. It is hopped to be small enough for implementation of this new approximants with neural networks.