Timely Split Inference in Wireless Networks: An Accuracy-Freshness Tradeoff

In this paper, we develop a novel framework that optimizes the tradeoff between the accuracy and freshness of an edge inference system, where a device extracts and transmits quantized feature vectors to an edge server for inference. We evaluate the inference accuracy in terms of discriminant gain, and measure the inference freshness by defining a metric termed age of edge inference (AoEI). To achieve timely and accurate edge inference, we minimize the average AoEI under an average inference distortion constraint by jointly optimizing the transmission and quantization policies. By using tools from Markov decision process, we develop an efficient policy to solve the problem. Simulations verify the accuracy-freshness tradeoff and the superiority of the proposed policy.