Learning Task-Oriented Communication for Edge Inference: An Information Bottleneck Approach

This paper investigates task-oriented communication for edge inference, where a low-end device transmits the extracted feature vector of local data sample to powerful server processing. It is critical encode into an informative and xmlns:xlink="http://www.w3.org/1999/xlink">compact representation low-latency inference given limited bandwidth. We propose learning-based scheme that jointly optimizes extraction, source coding, channel coding in manner, i.e., targeting downstream task rather than reconstruction. Specifically, we leverage information bottleneck (IB) framework formalize rate-distortion tradeoff between informativeness encoded performance. As IB optimization computationally prohibitive high-dimensional data, adopt variational approximation, namely (VIB), build tractable upper bound. To reduce overhead, sparsity-inducing distribution as prior VIB sparsify vector. Furthermore, considering dynamic conditions practical systems, variable-length encoding based on neural networks adaptively adjust activated dimensions different conditions. Extensive experiments evidence proposed system achieves better baseline methods significantly reduces transmission latency