Distributed data analytics at the network edge : an integrated communication-computing approach

Abstract

The evolution of mobile networks has led to a highly intelligent network, with unprecedented mobile data traffic driving the movement of data analytics to the edge of the network. This has given rise to distributed data analytics, where a central server orchestrates edge devices to execute collaborative data processing and knowledge distillation. These services provide edge intelligence and distributed sensing for various sectors, from e-healthcare to autonomous driving to IoT sensing. Advancing distributed data analytics relies on efficient integrated communication and computing (IC2) designs that execute communication techniques and computing algorithms coordinately, promising scalability, robustness, and efficiency. This dissertation explores IC2 designs to support particular applications in distributed data analytics. Its beginning focus is on leveraging IC2 within traditional one-shot architectures, which perform communication and computing sequentially. In the one-shot architecture, this dissertation first enhances distributed principal component analysis (DPCA) in wireless systems through a framework featuring analog multiple-input-multiple-output (MIMO) transmission. This approach utilizes uncoded analog transmission of local subspaces to estimate a global subspace, addressing channel distortion with maximum-likelihood subspace estimate- tors. The dissertation derives tight bounds on the estimation error, demonstrating that channel information can be dispensable for DPCA acceleration. Simulation results validate the theoretical findings, highlighting the promising latency performance of the proposed analog MIMO system. The second contribution addresses multi-view feature fusion in distributed sensing under the split inference architecture. By efficiently aggregating multiple sensor views using over-the-air computation (AirComp), this work establishes an analytical framework to quantify the fundamental performance gains from view- and-channel aggregation. The framework proves that end-to-end (E2E) sensing uncertainty diminishes exponentially as the number of sensor views increases, with performance validated through experiments using a convolutional neural network model and real-world datasets. Beyond traditional one-shot architectures, this dissertation introduces the FlyCom2 framework. It decomposes complex communication-and-computing tasks into low-complexity, parallel operations, thereby reducing on-device computation, halving E2E latency, and allowing continuous performance improvement. Relevant topics studied in this dissertation include distributed tensor decomposition (DTD) and representation learning for distributed Point Cloud (PtCloud) fusion, demonstrating significant performance gains and reduced computational complexity. In DTD scenarios, FlyCom2 is implemented to handle eigenspace extraction of high-dimensional data distributed across edge devices. The proposed framework enables streaming on-device computation with low complexity by leveraging a random sketching technique and achieves progressive global aggregation through the integration of progressive uploading and MIMO AirComp. A global sub-space estimator is designed to take accumulated observations to generate online estimates while suppressing AirComp-induced distortions. Its performance is evaluated by theoretical performance analysis with the obtained insights verified by experiments.

As for the PtCloud-based distributed sensing, FlyCom2 addresses the computational and communication challenges in PtCloud fusion. Underpinning the proposed framework is to align the PtCloud fusion with the process of Gaussian process regression. Then, the E2E fusion performance is optimized via a joint design of local observation synthesis and AirComp, which resolves the issues of communication distortions, data heterogeneity, and temporal correlation of FlyCom2. Validation on real-world datasets showcases the superiority of the proposed FlyCom2.