A Built-in Methodology for Resemblance Gathering in RKII Networks

Abstract

This paper presents a methodology to test RKII cells for dynamic resemblance.The cells are the basic processing blocks of RKII networks, which, as with other Artificial Neural Networks, are made of repeated processing elements interconnected in some pre-defined manner. For the RKII network, each processing element, the RKII, is a dynamic piece that behaves as an input controlled oscillator, therefore the network represents a set of coupled oscillators. Each RKII cell in the network should exhibit similar characteristics for suitable operation. However, in a real analogue CMOS VLSI implementation, similarity will change along the integrated circuit, due to process variations, failures, or even performance degradation. The variations found in different cells may prevent the network to operate properly. The present work presents a method to find and select a set of RKII cells, within the chip universe, that reflects a pre-defined degree of similitude. The method employs an iterative procedure that searches the network to find the set of cells that fit within a percentage of dynamic variation, and finds a maximum number of cells that best resemble each other. In the end, the most similar cells are selected while the others are turned off.