Learning at the network edge : quantization, multiple access, and constellation recognition

Abstract

To support intelligent applications such as autonomous driving and meet the demand on intelligent radios for accommodating diversified service requirements, recent years have witnessed the development of an intelligent network edge (INE), referring to the deployment of machine learning algorithms at edge servers. Depending on the roles of learning algorithms, the development of INE can be categorized into two paradigms, namely 1) communication-assisted learning and 2) learning-assisted communication. The former targets fast and accurate learning by developing enabling communication techniques while the latter constitutes a set of learning algorithms for automating communication systems. Advancing the two paradigms requires the seamless integration of two areas: wireless communication and machine learning. This dissertation contributes to the emerging area of INE by designing learning-oriented communication techniques and communication-oriented learning algorithms for both paradigms. For communication-assisted learning, novel communication techniques are developed to enable communication-efficient federated edge learning (FEEL). Specifically, to reduce the communication overhead incurred by high-dimensional stochastic gradient uploading, we first propose a hierarchical framework for quantizing the stochastic gradients. Underpinning the design is an innovative idea of decomposition of gradients into its norm, normalized block gradients, and the hinge vector, and their efficient quantization using three low-dimensional quantizers. Particularly, the decomposition allows the exploitation of Grassmannian codebooks to efficiently compress the gradients. Then, a quantization bit-allocation strategy is developed by dividing the total bits for quantizing each gradient to determine the resolutions of quantizers in the proposed framework under the criterion of minimum distortion, followed by learning convergence analysis. Next, to further accelerate FEEL, a novel multi-access scheme, coined as one-bit broadband digital aggregation (OBDA), is proposed. The new scheme features one-bit gradient quantization followed by digital quadrature amplitude modulation (QAM) at the edge devices and a majority-voting based decoding at the edge server. To understand the convergence performance of the proposed OBDA, a comprehensive analysis framework has been developed by accounting for the effects of wireless channel hostilities. For learning-assisted communication, we have developed a data clustering enabled approach for non-coherent modulation recognition so as to automate the communication system. The approach features an analytical framework comprising two key innovations. First, the equivalence between the expectation-maximization (EM) algorithm and K-means algorithm originally established in the Euclidean space is rediscovered on the Grassmann manifold. Second, theoretic analysis has been conducted to understand the performance of the designed learning algorithms by accounting for the effects of various parameters ranging from the signal-to-noise ratio to constellation size. Besides the framework, a novel symbol-bit-mapping scheme in a Grassmann codebook is proposed for enabling the simultaneous symbol-and-bit detection at the receiver.