Task-Oriented Communication for Multidevice Cooperative Edge Inference

This paper investigates task-oriented communication for multi-device cooperative edge inference, where a group of distributed low-end devices transmit the extracted features local samples to powerful server inference. While inference can overcome limited sensing capability single device, it substantially increases overhead and may incur excessive latency. To enable low-latency we propose learning-based scheme that optimizes feature extraction encoding in manner, i.e., remove data redundancy information is essential downstream task rather than reconstructing at server. Specifically, leverage Tishby’s bottleneck (IB) principle (Tishby et al., 2000) extract task-relevant each adopt (DIB) framework Aguerri Zaidi, 2021, formalize single-letter characterization optimal rate-relevance tradeoff encoding. admit flexible control overhead, extend DIB deterministic (DDIB) objective explicitly incorporates representational costs encoded features. As IB-based objectives are computationally prohibitive high-dimensional data, variational approximations make optimization problems tractable. compensate potential performance loss due approximations, also develop selective retransmission (SR) mechanism identify among multiple attain additional reduction. Extensive experiments on multi-view image classification object recognition tasks evidence proposed achieves better existing methods.