Multi-device edge intelligence : resources allocation, multiple access, and sensor selection

Abstract

The sixth-generation (6G) mobile networks are envisioned to realize ubiquitous artificial intelligence (AI) and support a wide range of AI-empowered applications, e.g., augmented reality and autonomous driving. Materializing this vision relies on an emerging paradigm, termed edge intelligence, which pushes AI towards the wireless edge in proximity to data sources to achieve low latency, high reliability, and enhanced privacy. Edge inference and distributed sensing are two representative use cases of edge intelligence. The former offloads computation-intensive inference tasks from mobile devices to edge servers, thereby enhancing device capabilities and extending battery life. The latter involves a fusion center acquiring and fusing sensory features from numerous edge devices for AI-empowered collaborative sensing tasks. However, conventional wireless systems designed under the rate-centric principle are inherently sub-optimal in terms of end-to-end (E2E) task performance. This dissertation addresses this challenge by designing task-oriented communication techniques for several key edge intelligence applications.

The dissertation first studies multi-device edge inference where an edge server handles inference requests from multiple users utilizing efficient inference techniques, batching and early exiting. A set of efficient algorithms are designed for joint allocation of communication-and-computation resources to maximize the inference throughput. For the case without early exiting, the target problem is solved optimally by nesting a threshold-based user selection scheme in a sequential search for the maximum batch size. For the general case with batching and early exiting, the dissertation further develops a low-complexity sub-optimal algorithm and an optimal algorithm based on depth-first tree-search with node pruning. Experimental results show that the proposed resource allocation scheme can double the inference throughput compared with conventional schemes.

Next, targeting distributed sensing, the dissertation proposes a novel framework, called Spatial Over-the-Air Fusion, which exploits radio waveform superposition to aggregate spatially sparse features over the air. The framework supports simultaneous aggregation over multiple voxels, which partition the sensing region, and across multiple subcarriers. It exploits both spatial feature sparsity and channel diversity to pair voxel-level aggregation tasks with subcarriers. The key challenge is to solve the resultant mixed-integer programming of Voxel-Carrier Pairing and Power Allocation (VoCa-PPA). An optimal algorithm is proposed through derivations of optimal power allocation as a closed-form function of voxel-carrier pairing and a useful property of VoCa-PPA that allows dramatic solution space reduction. Simulations using real datasets show significant improvements in sensing accuracy achieved by optimized VoCa-PPA.

Finally, this dissertation develops semantic-relevance-aware sensor selection to achieve optimal E2E task performance under heterogeneous sensor relevance and channel states. E2E sensing accuracy analysis is provided to characterize the sensing task performance in terms of selected sensors' relevance scores and channel states. Subsequently, the sensor-selection problem for accuracy maximization is formulated as an integer program and solved through a tight approximation of the objective. The near-optimal solution exhibits a priority-based structure, which ranks sensors based on a priority indicator combining relevance scores and channel states and selects top-ranked sensors. Experimental results on both synthetic and real datasets show substantial accuracy gain achieved by the proposed selection scheme compared to existing benchmarks.