EQ-Net: A Unified Deep Learning Framework for Log-Likelihood Ratio Estimation and Quantization

  • 2021-05-03 17:27:05
  • Marius Arvinte, Ahmed H. Tewfik, Sriram Vishwanath
  • 0

Abstract

In this work, we introduce EQ-Net: the first holistic framework that solvesboth the tasks of log-likelihood ratio (LLR) estimation and quantization usinga data-driven method. We motivate our approach with theoretical insights on twopractical estimation algorithms at the ends of the complexity spectrum andreveal a connection between the complexity of an algorithm and the informationbottleneck method: simpler algorithms admit smaller bottlenecks whenrepresenting their solution. This motivates us to propose a two-stage algorithmthat uses LLR compression as a pretext task for estimation and is focused onlow-latency, high-performance implementations via deep neural networks. Wecarry out extensive experimental evaluation and demonstrate that our singlearchitecture achieves state-of-the-art results on both tasks when compared toprevious methods, with gains in quantization efficiency as high as $20\%$ andreduced estimation latency by up to $60\%$ when measured on general purpose andgraphical processing units (GPU). In particular, our approach reduces the GPUinference latency by more than two times in several multiple-inputmultiple-output (MIMO) configurations. Finally, we demonstrate that our schemeis robust to distributional shifts and retains a significant part of itsperformance when evaluated on 5G channel models, as well as channel estimationerrors.

 

Quick Read (beta)

loading the full paper ...