Quantization Effects in a CNN-based Channel Estimator

Abstract

In this paper, we study the impact of the convolutional neural networks (CNN) quantization for the channel estimation. In the wireless network edge, with the adoption of deep learning (DL) algorithms, the limited computational resources bottleneck needs to be considered. Thus, a study using a field-programmable gate array (FPGA) platform is carried out, where the resource utilization and the timing requirements are analyzed. A single-input single-output orthogonal frequency-division multiplexing (OFDM) end-to-end link is adopted in this work. The bit error rate (BER) measures the quantization impact of the CNN-based channel estimation on the global system. The obtained results show that an improvement in the maximum operating frequency and in the resource efficiency can be obtained without deteriorating the end-to-end performance.