Communication-Efficient Vertically Split Inference via Over-the-Air Computation

In this paper, a vertically split neural network based edge-device collaborative inference framework is proposed to deal with the issue that heterogeneous raw data samples are obtained by different devices for an inference task as well as to enhance the feature extraction capability of edge devices. To alleviate the communication overhead caused by transmitting the high-dimensional local feature maps, the technique of over-the-air computation (AirComp) is adopted. Furthermore, a broadband channel is considered and orthogonal frequency division multiplexing (OFDM) is leveraged to support the simultaneous aggregation of all dimensions of the local results. Due to the channel fading and noise, a scheme of joint subcarrier allocation, power allocation, and receive beamforming is then proposed to minimize the aggregation distortion and improve the inference accuracy. Extensive experiments are conducted to verify the superiority of the proposed design over benchmarking schemes.