Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Collaborative Edge Inference via Semantic Grouping under Wireless Channel Constraints

Mateus P. Mota, Mattia Merluzzi, Emilio Calvanese Strinati

TL;DR

This work studies collaborative edge inference (edge ensembles) where devices exchange intermediate features rather than raw data to improve classification under wireless bandwidth constraints. It introduces a semantic grouping mechanism based on attention, using a key-query exchange to select collaborators and a pruning threshold to control communication; the model is trained with channel effects incorporated, enhancing robustness to wireless noise. Key findings show that the splitting point and communication pruning significantly impact accuracy vs. resource usage, and that the query channel requires higher reliability than the data channel, though its small size mitigates cost. The approach delivers robust edge inference with reduced communication overhead, offering practical guidance for scalable wireless-edge AI deployments and suggesting directions for extending channel-aware matching and query-guided data transmission across tasks.

Abstract

In this paper, we study the framework of collaborative inference, or edge ensembles. This framework enables multiple edge devices to improve classification accuracy by exchanging intermediate features rather than raw observations. However, efficient communication strategies are essential to balance accuracy and bandwidth limitations. Building upon a key-query mechanism for selective information exchange, this work extends collaborative inference by studying the impact of channel noise in feature communication, the choice of intermediate collaboration points, and the communication-accuracy trade-off across tasks. By analyzing how different collaboration points affect performance and exploring communication pruning, we show that it is possible to optimize accuracy while minimizing resource usage. We show that the intermediate collaboration approach is robust to channel errors and that the query transmission needs a higher degree of reliability than the data transmission itself.

Collaborative Edge Inference via Semantic Grouping under Wireless Channel Constraints

TL;DR

This work studies collaborative edge inference (edge ensembles) where devices exchange intermediate features rather than raw data to improve classification under wireless bandwidth constraints. It introduces a semantic grouping mechanism based on attention, using a key-query exchange to select collaborators and a pruning threshold to control communication; the model is trained with channel effects incorporated, enhancing robustness to wireless noise. Key findings show that the splitting point and communication pruning significantly impact accuracy vs. resource usage, and that the query channel requires higher reliability than the data channel, though its small size mitigates cost. The approach delivers robust edge inference with reduced communication overhead, offering practical guidance for scalable wireless-edge AI deployments and suggesting directions for extending channel-aware matching and query-guided data transmission across tasks.

Abstract

In this paper, we study the framework of collaborative inference, or edge ensembles. This framework enables multiple edge devices to improve classification accuracy by exchanging intermediate features rather than raw observations. However, efficient communication strategies are essential to balance accuracy and bandwidth limitations. Building upon a key-query mechanism for selective information exchange, this work extends collaborative inference by studying the impact of channel noise in feature communication, the choice of intermediate collaboration points, and the communication-accuracy trade-off across tasks. By analyzing how different collaboration points affect performance and exploring communication pruning, we show that it is possible to optimize accuracy while minimizing resource usage. We show that the intermediate collaboration approach is robust to channel errors and that the query transmission needs a higher degree of reliability than the data transmission itself.

Paper Structure

This paper contains 11 sections, 7 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. System Model
  3. Communication System
  4. Key Performance Indicators
  5. Semantic matching-based grouping
  6. Experiments and Discussion
  7. Simulation details and parameters
  8. Splitting point choice
  9. Channel effect on query and data transmission
  10. Communication pruning effect
  11. Conclusions and Perspectives

Figures (4)

  • Figure 1: System model scheme for the proposed collaborative inference problem.
  • Figure 2: Accuracy and # of features when varying the splitting point.
  • Figure 3: Accuracy for different query and intermediate data . itting point: 11.
  • Figure 4: Studying the effect of the communication pruning. The lowest thresholds are $\left[ 0, 0.001, 0.01 \right]$. Splitting Point: 11