Toward robust, reliable, and ultra-low-latency edge intelligence

Abstract

The rapid advancement of chip-level computing capabilities, coupled with the exponential proliferation of Internet-of-Things (IoT) devices, has ushered in an era of data explosion in wireless networks. This trend facilitates the migration of artificial intelligence (AI) from centralized cloud infrastructures to the network edge, closer to distributed data, giving rise to a new research area known as edge intelligence. One of the key features of edge intelligence is edge learning, which extracts AI from local data using machine learning algorithms; the other is integrated sensing and edge AI (ISEA), focusing on the efficient deployment of AI models at edge networks for environmental perception. However, the implementation of these functions faces two primary challenges: 1) interference-induced perturbations during data exchange between the edge server and devices, and 2) end-to-end (E2E) latency in task execution, involving data sensing, on-device computation, and communication. To address these challenges, this dissertation explores task-oriented designs to support robust, reliable, and ultra-low-latency edge intelligence.

To achieve robust edge intelligence, this dissertation proposes the spectrum breathing framework. This framework integrates data compression and spread-spectrum techniques to suppress interference without bandwidth expansion. The former operation contracts the data bandwidth for the subsequent spectrum spreading. Relevant topics explored in this dissertation include edge learning and distributed sensing. For robust edge learning, the proposed framework cascades the gradient pruning and spread spectrum for interference suppression. The convergence speed under this transceiver is characterized based on the supermartingale process, revealing the performance tradeoff between the pruning error and induced interference under constrained bandwidth. Optimization of the tradeoff yields two schemes that can be either fixed or adaptive to channels and the learning process, verified experimentally to achieve near-optimal performance. As for robust distributed sensing, spectrum breathing cascades feature compression and spread spectrum prior to feature transmission. With the metric of sensing accuracy, the tradeoff between these two operations is uncovered under a bandwidth constraint. The optimal feature compression strategy, adaptive to varying channel conditions, is derived by approximating sensing accuracy with a unimodal surrogate function. Evaluations on real-world datasets demonstrate the superiority of the proposed approach.

Targeting mission-critical and time-sensitive tasks, this dissertation achieves reliable and ultra-low-latency edge intelligence. One contribution enhances the reliability of latency-constrained edge inference systems by introducing and mathematically characterizing a goal-oriented reliability metric, termed inference outage (InfOut) probability. Constrained by E2E latency, this framework establishes a fundamental tradeoff between communication overhead and InfOut probability. This dissertation then enables tractable tradeoff optimization by deriving accurate surrogate functions for InfOut probability. Experiments demonstrate the significant performance gain from conventional communication-centric approaches. To meet strict latency requirements, the other contribution proposes an ultra-low-latency (ultra-LoLa) inference framework that integrates short-packet transmission with edge inference in distributed sensing. An efficient optimization approach is developed by characterizing the tradeoff between packet length and sensing observations. Experiments show its superiority over conventional reliability-oriented protocols.