Designing of array neuron-equivalentors with a quasi-universal activation function for creating a self-learning equivalent- convolutional neural structures

Abstract

In the paper, we consider the urgent need to create highly efficient hardware accelerators for machine learning algorithms, including convolutional and deep neural networks, for associative memory models, clustering, and pattern recognition. We show a brief overview of our related works the advantages of the equivalent models (EM) for designing bio-inspired systems. Such EM-paradigms are very perspective for processing, clustering, recognition, storing large size, strongly correlated, highly noised images and creating of uncontrolled learning machine. And since the basic nodes of EM are such vector-matrix (matrix-tensor procedures with continuous-logical operations as: normalized vector operations "equivalence", "nonequivalence", and etc., we consider in this paper new conceptual approaches to the design of full-scale arrays of such neuron-equivalentors (NEs) with extended functionality, including different activation functions. Our approach is based on the use of analog and mixed (with special coding) methods for implementing the required operations, building NEs (with number of synapsis from 8 up to 128 and more) and their base cells, nodes based on photosensitive elements and current mirrors. Simulation results show that the efficiency of NEs relative to the energy intensity is estimated at a value of not less than 10 12 an. op. / sec on W and can be increased. The results confirm the correctness of the possibility of creating NE and MIMO structures on their basis.