The Vector Decomposition Method

Abstract

When implementing neural networks in either software or hardware, one must have specifications for building blocks (either software functions or hardware circuits) in order to make a sufficiently good implementation. In this context, sufficiently good means that the minimum requirements for proper operation must be satisfied and that at the same time the various building blocks must not be over-specified as this is usually expensive (economically or computationally). To derive specifications for the various building blocks for neural networks, a number of ways can be followed. We list three possible approaches and give a short discussion of all.