Tiny Machine Learning (Tiny-ML) for Efficient Channel Estimation and Signal Detection
Hongfu Liu, Ziping Wei, Hengsheng Zhang, Bin Li, Chenglin Zhao
Abstract
Machine learning has provided great potential in intelligent signal processing, e.g., channel estimation and signal detection. But still, it is difficult to deploy deep neural network (DNN) on low-cost and low-power edge devices, e.g., in both training and inference phases. For one thing, the training of massive network parameters incurs a huge computational complexity and rarely becomes feasible; and for another, even a trained DNN involves still high computations and demands considerable storage size. In this work, we present a tiny machine learning (Tiny-ML) approach for hardware-efficient channel estimation and signal detection. The key innovation of our Tiny-ML is that we replace each large dense layer of DNN with three small cascading sub-layers; and therefore, the computation/storage of a large matrix is replaced with that of small ones. To enable the lightweight training of our Tiny-ML, a novel rank-restricted back-propagation algorithm is further designed. Numerical simulations demonstrate the advantages of our new method. Without sacrificing the estimation and detection accuracy, our Tiny-ML attains an acceleration of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\rm {\mathbf {2.5\sim\!3\times }}$</tex-math></inline-formula> in model training and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\rm {\mathbf {\sim\!4.5\times }}$</tex-math></inline-formula> in model inference, as well as a substantial storage reduction of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\rm {\mathbf {\sim\!4.5\times }}$</tex-math></inline-formula> , when compared to classical fully-connected (FC) DNN. As such, our method paves the way for deploying tiny AI onto low-cost and low-power hardware devices, thereby exploiting the full potential of intelligent signal processing.