A Simplified Realization for Data-Driven MIMO Detector With Product Quantization

Abstract

As research progresses, an increasing number of deep neural networks (DNNs) are applied to the multiple-input multiple-output (MIMO) detectors. This letter presents a simplified realization of the multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs) in data-driven MIMO detectors, which are implemented in the ChannelNet detector. ChannelNet is a data-driven MIMO detector exclusively crafted from readily available DNN components, such as MLPs and CNNs. The implementation of the MLP and CNN can be converted to matrix multiplication and replaced by a series of simple operations involving comparisons, querying the lookup tables, and integer additions. By setting a target loss in advance, a sequentially iterative simplification of ChannelNet is achieved. Numerical results demonstrate that ChannelNet has superior performance, and the simplified ChannelNet reduces at least half of the computational complexity with less than 0.2 dB performance loss at the expense of acceptable storage space.