CMOS Implementation of the Trainee’s Threshold Logical Element. Part I. Design and Training Diagram
Keywords:
Artificial Neuron, Synapse, Threshold Logical Learning Element, Training Algorithm, Learning Step, CMOS Technology, Threshold Logical FunctionAbstract
Purpose: The objective is to show a possibility of implementation an analog-digital artificial neuron on the example of building
a logical threshold element learning complex logical threshold functions in CMOS technology which uses modern design standards.
Methods: representation of McCulloch — Pitts neuron in the form of relation of the total of weighted inputs to the threshold and development
of a methodology of designing a threshold logical learning element consisting of synopses which conductivity depends on
input variables and their threshold weights reduced to function which are accumulated during a learning process in analog memory
elements; а high sensitive comparator which compares total conductivity of synapses with conductivity of its p-channel part represents
the highest function threshold value; and three output amplifiers with different firing thresholds. Results: It has been shown that
implementability of a threshold learning element depends only on a function threshold value and does not depend on the total of input
weights and their number. The element can be trained to implement an arbitrary threshold function which threshold does not exceed
a given value. The element circuit considered in the paper is oriented towards the maximum threshold value equal to 89 and is capable
to implement any threshold function of 10 variables. There has been proposed a training diagram which provides parallel forming of
weights for active inputs and makes an automatic choice a value of a learning step. All practical results are received using PSPICE
simulation of circuits constructed in CMOS technology of 0.18 micron. Practical relevance: There have been considerably extended
functional possibilities of the proposed threshold learning element. It can be applied in logical systems of image recognition and for
creation a new generation of neuron chips.