Toward Real-Time Edge AI: Model-Agnostic Task-Oriented Communication with Visual Feature Alignment

TL;DR

The paper tackles cross-model task-oriented communication for real-time edge AI by addressing the problem of incoherent feature spaces across independently trained TOC systems. It introduces anchor data as a shared knowledge base and proposes two alignment paradigms: server-based feature alignment leveraging linear invariance with estimators (MLP, LS, MMSE) to learn a transformation between feature spaces, and on-device alignment using relative representations based on angle-preserving properties and cosine similarities. The authors provide theoretical insights, including a Lipschitz-based bound on alignment error (Proposition 1) and practical training/estimation strategies, and validate the approach on SVHN and CIFAR-10 across multiple architectures and channel conditions. Experiments demonstrate significant gains in cross-model inference accuracy with low latency and memory overhead, supporting real-time applicability and model-agnostic cross-provider edge inference. The work advances cross-provider edge AI by enabling interoperable, efficient, and scalable task-oriented communications at the network edge.

Abstract

Task-oriented communication presents a promising approach to improve the communication efficiency of edge inference systems by optimizing learning-based modules to extract and transmit relevant task information. However, real-time applications face practical challenges, such as incomplete coverage and potential malfunctions of edge servers. This situation necessitates cross-model communication between different inference systems, enabling edge devices from one service provider to collaborate effectively with edge servers from another. Independent optimization of diverse edge systems often leads to incoherent feature spaces, which hinders the cross-model inference for existing task-oriented communication. To facilitate and achieve effective cross-model task-oriented communication, this study introduces a novel framework that utilizes shared anchor data across diverse systems. This approach addresses the challenge of feature alignment in both server-based and on-device scenarios. In particular, by leveraging the linear invariance of visual features, we propose efficient server-based feature alignment techniques to estimate linear transformations using encoded anchor data features. For on-device alignment, we exploit the angle-preserving nature of visual features and propose to encode relative representations with anchor data to streamline cross-model communication without additional alignment procedures during the inference. The experimental results on computer vision benchmarks demonstrate the superior performance of the proposed feature alignment approaches in cross-model task-oriented communications. The runtime and computation overhead analysis further confirm the effectiveness of the proposed feature alignment approaches in real-time applications.

Figures (12)

• Figure 1: The illustration of the compatibility issue in cross-model task-oriented communications.
• Figure 2: The independent task-oriented communication models, denoted as TOC-1 and TOC-2, and the anchor data shared between them. The parameters $(\boldsymbol{\phi}_1, \boldsymbol{\theta}_1)$ and $(\boldsymbol{\phi}_2, \boldsymbol{\theta}_2)$ for each task-oriented communication model are trained with different architectures and independent training processes. Anchor data works as a common knowledge base and all the task-oriented communication models can recognize its data samples and data index.
• Figure 3: Probabilistic modeling for cross-model task-oriented communication.
• Figure 4: The server-based feature alignment using single-layer MLP for cross-model task-oriented communications.
• Figure 5: The proposed on-device feature alignment using relative representation encoding.

Theorems & Definitions (3)

• Definition 1
• Proposition 1
• proof