Integrated Sensing-Communication-Computation for Over-the-Air Edge AI Inference

Zeming Zhuang, Dingzhu Wen, Yuanming Shi, Guangxu Zhu, Sheng Wu, Dusit Niyato

TL;DR

This paper introduces a task-oriented framework for Integrated Sensing-Communication-Computation (ISCC) to boost the performance of Edge AI. By jointly optimizing how devices sense the environment and transmit data using Over-the-Air Computation (AirComp), the authors achieve significantly higher and more balanced classification accuracy, specifically by focusing on the "hardest-to-distinguish" pairs of data classes.

What Problem Does This Paper Address?

In the "Intelligence of Everything" (IoE) era, edge devices (like sensors or autonomous cars) often lack the power to run massive AI models. Edge-device co-inference solves this by splitting the work: devices extract features, and a server completes the inference.

However, traditional systems treat sensing, communication, and computation as separate silos. Standard communication metrics like Signal-to-Noise Ratio (SNR) or bit rate don't necessarily correlate with how well an AI model can actually classify an object. Furthermore, when multiple devices sense the same target (wide-view sensing), they produce redundant, noise-corrupted data. The challenge is: how do we jointly manage power and bandwidth across these three stages to maximize the AI's "understanding" rather than just the data throughput?

Methods

The authors propose a task-oriented ISCC scheme with several novel components:

  • Maximum Minimum Pair-wise Discriminant Gain: Instead of maximizing average accuracy (which can leave some classes poorly separated), the authors use a "max-min" approach. They maximize the separation (Kullback-Leibler divergence) of the two classes that are most likely to be confused. This leads to a more balanced and robust model.
  • AirComp (Over-the-Air Computation): To handle the communication bottleneck, the system uses AirComp over OFDM. Instead of devices taking turns to send data, they transmit simultaneously on the same frequency. The wireless channel itself performs a "weighted summation" of the features, making the communication overhead independent of the number of devices.
  • Joint Optimization: The paper formulates a complex optimization problem involving sensing power assignment, transmit precoding, and receive beamforming. They solve this non-convex problem using Successive Convex Approximation (SCA) and a primal-dual method.
  • Feature Extraction: They utilize Singular Value Decomposition (SVD) for clutter cancellation and Principal Component Analysis (PCA) for low-dimensional feature extraction at the device level.

Datasets

The researchers evaluated their framework using a Human Motion Recognition task.

  • Source: Data was generated using a specialized wireless sensing simulator.
  • Classes: Four distinct motions—adult pacing, adult walking, child pacing, and child walking.
  • Scale: A training dataset of 6,400 samples (ground-truth/noise-free) and a test dataset of 1,600 samples (distorted by simulated clutter and noise).

Key Insights and Ideas

  • Task-Oriented vs. Data-Oriented: The "Aha!" moment is that communication resources should be allocated based on the importance of a feature to the final AI decision, not just the raw signal strength.
  • The "Weakest Link" Strategy: By focusing on the minimum discriminant gain, the system ensures that the most similar classes (e.g., "adult pacing" vs. "child pacing") are pushed apart in the feature space, which raises the floor for overall system accuracy.
  • Sensing-Comm Coupling: The quality of the features extracted (computation) is directly limited by the sensing power, and the preservation of those features is limited by the communication channel. Optimizing these together provides "extra degrees of freedom" that sequential optimization lacks.

Key Results

  • Accuracy Boost: The proposed scheme significantly outperformed baselines. For example, in a Support Vector Machine (SVM) model, the proposed scheme reached nearly 95% accuracy as the discriminant gain increased, while a baseline focusing on average gain resulted in much lower accuracy for difficult class pairs.
  • Scalability: Thanks to AirComp, the inference accuracy increases as more devices are added (due to better noise suppression through aggregation) without requiring more bandwidth.
  • Energy Efficiency: The joint optimization allows the system to maintain high accuracy even when device energy thresholds are low, by intelligently shifting power between sensing and transmission.

Limitations and Future Work

  • Static Channels: The current model assumes "static" wireless channels during the transmission of a feature vector. Real-world channels are often time-variant, which would require more complex tracking.
  • Device Scheduling: The paper assumes all devices participate. In massive IoT networks, the server might need to select a subset of the "best" devices to save total network energy.
  • Partial Views: The study focuses on "wide-view" sensing where everyone sees the same thing. Future work could explore "partial views," where different devices see different parts of a target (e.g., one sees the front of a car, another sees the side).

Summary updated

Abstract

Edge-device co-inference refers to deploying well-trained artificial intelligent (AI) models at the network edge under the cooperation of devices and edge servers for providing ambient intelligent services. For enhancing the utilization of limited network resources in edge-device co-inference tasks from a systematic view, we propose a task-oriented scheme of integrated sensing, computation and communication (ISCC) in this work. In this system, all devices sense a target from the same wide view to obtain homogeneous noise-corrupted sensory data, from which the local feature vectors are extracted. All local feature vectors are aggregated at the server using over-the-air computation (AirComp) in a broadband channel with the orthogonal-frequency-division-multiplexing technique for suppressing the sensing and channel noise. The aggregated denoised global feature vector is further input to a server-side AI model for completing the downstream inference task. A novel task-oriented design criterion, called maximum minimum pair-wise discriminant gain, is adopted for classification tasks. It extends the distance of the closest class pair in the feature space, leading to a balanced and enhanced inference accuracy. Under this criterion, a problem of joint sensing power assignment, transmit precoding and receive beamforming is formulated. The challenge lies in three aspects: the coupling between sensing and AirComp, the joint optimization of all feature dimensions’ AirComp aggregation over a broadband channel, and the complicated form of the maximum minimum pair-wise discriminant gain. To solve this problem, a task-oriented ISCC scheme with AirComp is proposed. Experiments based on a human motion recognition task are conducted to verify the advantages of the proposed scheme over the existing scheme and a baseline.
BibTeX
@article{zhuang2024integrated,
  author = {Zeming Zhuang and Dingzhu Wen and Yuanming Shi and Guangxu Zhu and Sheng Wu and Dusit Niyato},
  title = {{Integrated Sensing-Communication-Computation for Over-the-Air Edge AI Inference}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2024},
  doi = {10.1109/twc.2023.3306465},
}
Figure