Timely Inference over Networks

Abstract

Next-generation communications (Next-G), such as 6G, are expected to support emerging networked intelligent systems, including autonomous driving, factory automation, digital twin technology, unmanned aerial vehicle (UAV) navigation, and extended reality. Timely inference is crucial for these networked intelligent systems. In this dissertation, we investigate a remote inference system, where a trained neural network is used to infer time-varying targets (e.g., the locations of vehicles and pedestrians) based on features (e.g., video frames) observed at a sensing node (e.g., a camera). The inference error is determined by (i) the timeliness and (ii) the sequence length of the features, where we use the Age of Information (AoI) as a metric for timeliness. In the first part of the dissertation, we discuss how to evaluate the importance of timely information in remote inference and the monotonicity of information aging. One might expect that the performance of a remote inference system degrades monotonically as the feature becomes stale. Using a new information-theoretic analysis, we show that this is true if the feature and target data sequence can be closely approximated as a Markov chain; it is not true if the data sequence is far from Markovian. Hence, the inference error is a function of the AoI, where the function could be non-monotonic. In addition, a longer feature can typically provide better inference performance, but it often requires more channel resources for sending the feature. In the second part of the dissertation, we study the transmission scheduling problem that optimizes timeliness and feature sequence length to minimize inference error. We introduce a new "selection-from-buffer" medium access model for status updating and minimize inference errors for both Markovian and non-Markovian data. For single-source and single-channel remote inference networks, we obtain optimal scheduling policies for both the cases of time-invariant and time-variant feature lengths. In multi-sources and multi-channel remote inference networks, the selection-from-buffer scheduling problem is a multi-action restless multi-arm bandit problem. For this setting, we design new scheduling policies by utilizing the Whittle index and duality-based gain index. The new scheduling policies are proven to be asymptotically optimal. Data-driven evaluations show that our policies can reduce inference errors by up to 10,000 times.