Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Revisiting Outage for Edge Inference Systems

Zhanwei Wang, Qunsong Zeng, Haotian Zheng, Kaibin Huang

TL;DR

The work extends reliability analysis to edge inference by defining InfOut, the end-to-end probability that inference accuracy falls below a threshold under latency constraints. It introduces a Gaussian surrogate for the receive discriminant gain, deriving a fundamental C^2 tradeoff: more observations and feature uploads improve inference reliability but incur higher latency. For linear classifiers, it proves a concave surrogate function of the number of transmitted features has a unique optimum, and extends the framework to CNNs by mapping CNN accuracy to a receive DG and applying a data-driven parameter estimation approach. Experiments on synthetic and real data show substantial gains over communication-centric baselines, demonstrating the practical impact of optimally balancing computation and communication in latency-constrained edge AI systems.

Abstract

One of the key missions of sixth-generation (6G) mobile networks is to deploy large-scale artificial intelligence (AI) models at the network edge to provide remote-inference services for edge devices. The resultant platform, known as edge inference, will support a wide range of Internet-of-Things applications, such as autonomous driving, industrial automation, and augmented reality. Given the mission-critical and time-sensitive nature of these tasks, it is essential to design edge inference systems that are both reliable and capable of meeting stringent end-to-end (E2E) latency constraints. Existing studies, which primarily focus on communication reliability as characterized by channel outage probability, may fail to guarantee E2E performance, specifically in terms of E2E inference accuracy and latency. To address this limitation, we propose a theoretical framework that introduces and mathematically characterizes the inference outage (InfOut) probability, which quantifies the likelihood that the E2E inference accuracy falls below a target threshold. Under an E2E latency constraint, this framework establishes a fundamental tradeoff between communication overhead (i.e., uploading more sensor observations) and inference reliability as quantified by the InfOut probability. To find a tractable way to optimize this tradeoff, we derive accurate surrogate functions for InfOut probability by applying a Gaussian approximation to the distribution of the received discriminant gain. Experimental results demonstrate the superiority of the proposed design over conventional communication-centric approaches in terms of E2E inference reliability.

Revisiting Outage for Edge Inference Systems

TL;DR

The work extends reliability analysis to edge inference by defining InfOut, the end-to-end probability that inference accuracy falls below a threshold under latency constraints. It introduces a Gaussian surrogate for the receive discriminant gain, deriving a fundamental C^2 tradeoff: more observations and feature uploads improve inference reliability but incur higher latency. For linear classifiers, it proves a concave surrogate function of the number of transmitted features has a unique optimum, and extends the framework to CNNs by mapping CNN accuracy to a receive DG and applying a data-driven parameter estimation approach. Experiments on synthetic and real data show substantial gains over communication-centric baselines, demonstrating the practical impact of optimally balancing computation and communication in latency-constrained edge AI systems.

Abstract

One of the key missions of sixth-generation (6G) mobile networks is to deploy large-scale artificial intelligence (AI) models at the network edge to provide remote-inference services for edge devices. The resultant platform, known as edge inference, will support a wide range of Internet-of-Things applications, such as autonomous driving, industrial automation, and augmented reality. Given the mission-critical and time-sensitive nature of these tasks, it is essential to design edge inference systems that are both reliable and capable of meeting stringent end-to-end (E2E) latency constraints. Existing studies, which primarily focus on communication reliability as characterized by channel outage probability, may fail to guarantee E2E performance, specifically in terms of E2E inference accuracy and latency. To address this limitation, we propose a theoretical framework that introduces and mathematically characterizes the inference outage (InfOut) probability, which quantifies the likelihood that the E2E inference accuracy falls below a target threshold. Under an E2E latency constraint, this framework establishes a fundamental tradeoff between communication overhead (i.e., uploading more sensor observations) and inference reliability as quantified by the InfOut probability. To find a tractable way to optimize this tradeoff, we derive accurate surrogate functions for InfOut probability by applying a Gaussian approximation to the distribution of the received discriminant gain. Experimental results demonstrate the superiority of the proposed design over conventional communication-centric approaches in terms of E2E inference reliability.

Paper Structure

This paper contains 31 sections, 4 theorems, 54 equations, 11 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Models and Metrics
  3. Sensing and Computation Model
  4. Data Distribution Model
  5. Communication Model
  6. Inference Model
  7. Linear Classification
  8. CNN Classification
  9. Relevant Metrics
  10. Inference Outage Probability
  11. On-device Feature Importance
  12. Analysis of Inference Outage Probability
  13. Tractable Surrogate of Inference Accuracy
  14. Inference Outage Probability
  15. Minimization of Inference Outage Probability
  16. ...and 16 more sections

Key Result

Lemma 1

The inference accuracy with $K$ observations and received feature set $\Tilde{\mathcal{S}}$, denoted as $a (K,\Tilde{\mathcal{S}})$, is lower bounded by where $G_{\sf R}$ is defined as the receive DG per observation: $\hat{W}_d$ is the minimum DG of the $d$-th feature dimension in eq:min_DG.

Figures (11)

  • Figure 1: The transceiver framework of the edge inference system.
  • Figure 2: The PDF comparison between the real distribution and Gaussian approximation. The simulation parameters are set as: $S=100, P_{\sf{act}}=0.8$.
  • Figure 3: InfOut probability under different channel outage probability. The numerical settings are $D=1000,L=2,A_{\sf{th}}=99.99\%$, $T_\Delta=10$$\mu$s, $T=10$ ms, $f_c=2.5$ GFLOPs/s, $N_F= 2.3$ MFLOPs.
  • Figure 4: An example of DG function with $D=4$.
  • Figure 5: Under the magnitude based feature selection scheme, the distribution comparison between inference accuracy and receive CNN DG using VGG16 model simonyan2015deep the ModelNet dataset ModelNet-Ref. The settings are $D=512, L=20, \alpha=1, \theta =1, K=12,P_{\sf{act}}=0.7$.
  • ...and 6 more figures

Theorems & Definitions (7)

  • Lemma 1: ZW2024ultra-LoLa
  • Lemma 2: Lindeberg's Condition feller1991introduction
  • Lemma 3: Distribution of Receive DG
  • Remark 1: Computation-communication tradeoff
  • Definition 1: Discriminant Gain Function
  • Proposition 1: Optimal Number of Selected Features
  • Definition 2: Receive CNN Discriminant Gain