From complex algorithms to analog signal processing: Generalized recurrent neural networks

Abstract

Approaching the information theoretic limits in a wireless communications system generally requires the application of complex algorithm. A class of algorithms facilitating this task are iterative methods such as the well-known iterative decoding of turbo codes or low density parity check codes. But also iterative equalization methods become more and more important in praxis. Usually, the signal processing for these iterative methods is realized with digital hardware. Besides all its advantages, a major drawback of digital signal processing is the large power consumption, especially for high data rates. The fundamental structure underlying iterative interference cancellation algorithms is a recurrent neural network (RNN). Recently, it has been shown that an iterative decoding algorithm based on belief propagation can be represented by a generalized RNN. Taking the original work of Hopfield as a starting point, we propose an analog electronic circuit which resembles a generalized RNN. Compared to digital signal processing, analog signal processing allows to perform iterative decoding or equalization with increased computational speed and reduced chip area and power consumption.