A Bit-Parallel Representation of Activation Functions for Fast Neural Networks

A procedure for constructing a neuron with fast calculation of the activation function by the tabular-algorithmic method is suggested. The effect is achieved by means of a bit-parallel representation of the result of computation of a nonlinear function oriented to the operation of group summation. The approach can be applied to activation functions of various forms, including logistical and sigmoid rational functions. The creation of a fast neuron requires hardware implementation of a nonlinear function on the basis of ROM, registers, hardware logic, and adders.