Recursive cellular nonlinear neural networks for ultra low noise digital arithmetic英文电子书.pdf

Recursive Cellular Nonlinear Neural Networks for Ultra-low Noise Digital Arithmetic J.J. Yeboah Jr., G.A. Jullien, J. W. Haslett ATIPS Laboratory, Dept. of Electrical and Computer Engineering University of Calgary, Alberta, Canada E-mail: {yeboah,jullien, haslett}@atips.ca Abstract - An innovative method of designing a class of analog 2. CELLULAR NEURAL NETWORKS cellular neural networks - Recursive Cellular Nonlinear (Neu- Recursive CNNs are based on the original concept of Cellular ral) Networks (RCNNs) - for ultra-low noise digital arithmetic Neural Networks as introduced by Leon Chua and Lin Yang. computation is presented. The intended application of the RCNN is for digital arithmetic in a sensitive mixed-signal envi- In [21 they described a cell composed of a linew capacitor, a linear resistor, and a nonlinear controlled currnt source, as ronment, such as a digital interface to a low output sensor, where digital noise is to be kept to a mininum. Our ultra low-noise shown in Figure 1. approach essentially employs asynchronous analog circuit con- cepts. In this paper we have analyzed and summarized the recursive arcitecture of our RCNN networks at the mathemati- cal and circuit level. An analog CMOS design of a 4-bit RCNN adder, which wil be used to verify the low-noise behaviour of our approach, is also presented . 1. INTRODUCTION There is great interest at present in the design of asynchro- nous systems based on the use of self-timing components for Figure 1 CNN Cell arithmetic. The increase in component density of monolit