Non-deterministic Quantization for mmWave Beam Prediction

In this paper, we propose a low complexity yet powerful deep neural network (DNN) that learns from binary quantified features for the millimeter wave (mmWave) beam prediction. The proposed model is capable of achieving a high reliable beam prediction and low hardware overhead, depending on the configurable DNN model in mmWave massive multipleinput multiple-output (massive MIMO) system. In order to overcome the noise and the mmWave channel fading attenuation, we firstly incorporate the ideas from non-deterministic quantization to extract a sparse feature representation. Using the nondeterministic method also allows the DNN model to compress the size of parameters, which enables the proposed model easy to implement. Secondly, matching the non-deterministic scheme with multiple quantizations, we build the static layers to extract an appropriate representation from binary sequence and infer by a wise scoring scheme based on voting results from multi-quantization. According to the experiment results, the accuracy can be improved from 77% to 94% compared with the conventional proposal under the 20 dB signal-to-noise ratio (SNR). And the proposed model can reduce the scale of parameters by 70%. We further accelerate the training time up to 34% by applying automatic mixed precision (AMP).