Design and simulation of cells, that realize arbitrary functions of activations of neurons in self-learning equivalent-convolutional neural structures

Abstract

We consider the urgent need to create hardware accelerators for CNN. We show a overview of the equivalent models (EMs), EM-paradigms for recognition images and learning with CL- operations as: "equivalence". We consider approaches to the design of arrays of neuron-equivalentors (NEs) with different activation functions. Approach is based on the use of mixed methods, building NEs (with number of synapsis 128) and their cells based current mirrors. Simulations show that the efficiency of NEs relative to the energy is estimated at a value of not less 10^12 an. op. / sec on W.