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. 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.
Introduction
In the era of the "Intelligence of Everything" (IoE), edge devices such as autonomous vehicles and IoT sensors are increasingly required to perform complex AI tasks. However, these devices often lack the computational power to run massive AI models locally. Edge-device co-inference addresses this by splitting the workload: devices perform initial feature extraction, while a powerful edge server completes the inference.
Traditional systems treat sensing, communication, and computation as independent silos. Standard communication metrics, such as Signal-to-Noise Ratio (SNR) or raw bit rate, often fail to correlate with the actual performance of an AI model in classification tasks. Furthermore, when multiple devices sense the same target (wide-view sensing), they produce redundant, noise-corrupted data. The challenge lies in jointly managing power and bandwidth across these three stages to maximize the AI's "understanding" rather than just data throughput.
System Architecture and Methodology
The proposed framework introduces a task-oriented Integrated Sensing-Communication-Computation (ISCC) scheme. As illustrated in the system architecture below, multiple devices sense a target simultaneously, extract features, and transmit them to an edge server.
Over-the-Air Computation (AirComp)
To overcome the communication bottleneck, the system utilizes AirComp over OFDM. Instead of devices taking turns to transmit data (which consumes significant bandwidth), they transmit simultaneously on the same frequency. The wireless channel itself performs a "weighted summation" of the features. This approach makes the communication overhead independent of the number of devices, significantly improving scalability.
Maximum Minimum Pair-wise Discriminant Gain
A core innovation of this work is the shift from maximizing average accuracy to a "max-min" approach. Instead of optimizing for the average separation between all classes, the authors maximize the separation (specifically the Kullback-Leibler divergence) of the two classes most likely to be confused.
By focusing on the "weakest link"—the most similar class pairs—the system ensures that the feature space is optimized for the hardest classification decisions, leading to more robust and balanced model performance.
Experimental Evaluation
The framework was evaluated using a Human Motion Recognition task involving four distinct motions: adult pacing, adult walking, child pacing, and child walking. The researchers used a training dataset of 6,400 noise-free samples and a test dataset of 1,600 samples distorted by simulated clutter and noise.
Key Results
- Accuracy Boost: The proposed scheme significantly outperformed baselines. Using a Support Vector Machine (SVM) model, the scheme reached nearly 95% accuracy as the discriminant gain increased. In contrast, baselines focusing on average gain resulted in much lower accuracy for difficult class pairs.
- Scalability: Thanks to AirComp, inference accuracy actually increases as more devices are added. This is due to better noise suppression through feature aggregation, achieved without requiring additional bandwidth.
- Energy Efficiency: Joint optimization allows the system to maintain high accuracy even under low energy thresholds by intelligently shifting power between the sensing and transmission stages.
Conclusion and Future Work
This paper demonstrates that communication resources should be allocated based on the importance of a feature to the final AI decision. By coupling sensing, communication, and computation, the system gains extra degrees of freedom that sequential optimization lacks.
Future research will address limitations such as static channel assumptions, exploring time-variant channels that are more common in real-world scenarios. Additionally, the team plans to investigate device scheduling in massive IoT networks and partial-view sensing, where different devices capture different perspectives of a single target.
Summary updated