Black-Box Edge AI Model Selection with Conformal Latency and Accuracy Guarantees

TL;DR

This paper introduces a black-box model selection framework for edge AI that guarantees latency and accuracy in real-time wireless environments, crucial for 6G applications like autonomous driving.

Contribution

It presents a novel conformal risk control-based method for selecting optimal models under latency and accuracy constraints in wireless edge AI.

Findings

Framework effectively meets deadline and accuracy requirements.

Dynamic model selection adapts to changing channel conditions.

Numerical validation on image classification demonstrates reliability.

Abstract

Edge artificial intelligence (AI) will be a central part of 6G, with powerful edge servers supporting devices in performing machine learning (ML) inference. However, it is challenging to deliver the latency and accuracy guarantees required by 6G applications, such as automated driving and robotics. This stems from the black-box nature of ML models, the complexities of the tasks, and the interplay between transmitted data quality, chosen inference model, and the random wireless channel. This paper proposes a novel black-box model selection framework for reliable real-time wireless edge AI designed to meet predefined requirements on both deadline violation probability and expected loss. Leveraging conformal risk control and non-parametric statistics, our framework intelligently selects the optimal model combination from a collection of black-box feature-extraction and inference models of varying complexities and computation times. We present both a fixed (relying on channel statistics) and a dynamic (channel-adaptive) model selection scheme. Numerical results validate the framework on a deadline-constrained image classification task while satisfying a maximum misclassification probability requirement. These results indicate that the proposed framework has the potential to provide reliable real-time edge AI services in 6G.