Inference at the wireless edge : capacity, progressive transmission, and multiple access

Abstract

Artificial intelligence (AI) is flourishing in wide-ranging practices around the globe. A massive number of users and terminals access AI services deployed in central clouds. It is trending to relocate AI inference deployment to edge networks for the desired distributed computing, low-latency, and privacy-preserving features. This future vision is termed edge inference, where edge devices and servers cooperate for inference computing. Despite its potential gains, edge inference is confronted with the critical communication bottlenecks. The bottlenecks are essentially raised from hostile wireless channels, high-dimensional feature data, and multiple access. Addressing these issues gives rise to one of the core missions of the sixth-generation (6G) network, namely advanced wireless access for edge inference. The conventional rate- and connectivity-centric design philosophy shall be shifted to 6G task-oriented networking. We materialize task-oriented wireless techniques tailored for edge inference in this dissertation.

To start with, we analyze the performance of the edge-inference systems specialized for classification. The communication throughput limits feature sizes in uploading. Classification accuracy varies among different tasks due to diversified feature dimensions and class discernibility. We propose a set of metrics under a terminology called classification capacity, referring to the maximum number of classes subject to communication and accuracy constraints. Leveraging the differential geometry tools allows us to quantify the impact of Gaussian and fading channels on subspace-classification performance in the proposed notion. In the asymptotic regime, the classification capacity with task selection exponentially increases as the communication rate grows for all considered channel configurations. In contrast, the lower bounds on classification capacity linearly scale with the communication rate in that regime if the server, more practically, randomly admits tasks.

To tackle high-dimensional data, we propose a progressive feature transmission protocol for edge inference. The proposed protocol progressively transmits prioritized features with early termination. It is materialized by two key modules termed importance-aware feature selection and transmission-termination control, respectively. The former prioritizes the features in descending order of their contributions to inference performance such as confidence and accuracy. The latter intelligently stops progressive transmission when further delivering features comes with marginal confidence gain against the incurred communication cost. This approach thereby boosts communication efficiency. Concrete designs are developed for linear classification and neural network models via solving the corresponding stochastic control problems, respectively.

On multiple access, we propose a task-oriented simultaneous access scheme for multi-view pooling in edge inference, called over-the-air multi-view pooling. It exploits wave superposition to allow all sensors to simultaneously transmit the local features and the fusion center to receive the over-the-air pooled feature. We design a pair of configurable pre- and post-processing functions to admit a broader range of pooling functions, including max-pooling. Our analysis quantifies its impact on inference accuracy and presents a trade-off between function approximation and noise suppression. The analysis yields simple yet effective configuration optimization for representative over-the-air average-pooling and max-pooling. Last, we extend over-the-air schemes to more practical two-cell multiple-input-multiple-output networks. The proposed simultaneous signal-and-interference alignment achieves computing degrees-of-freedom up to the half of the antenna array size.