Learning and Communications Co-Design for Remote Inference Systems: Feature Length Selection and Transmission Scheduling

TL;DR

This work tackles remote inference by jointly optimizing feature length and transmission scheduling under AoI constraints, addressing both single- and multi-source scenarios. It develops information-theoretic foundations linking feature length and AoI to inference error, derives a closed-form threshold policy for time-invariant feature length, and extends to time-variant length via a tailored SMDP policy, with policy-iteration-based computation of optimal cost and value functions. For multiple sources, it employs Lagrangian relaxation to decompose the problem into per-source MDPs and introduces the Net Gain Maximization policy that avoids indexability requirements, solving a knapsack-like allocation each slot. Trace-driven results show dramatic gains—up to about $10^4$× in single-source and substantial improvements in multi-source cases—highlighting the practical impact for remote AI-enabled systems with constrained channels.

Abstract

In this paper, we consider a remote inference system, where a neural network is used to infer a time-varying target (e.g., robot movement), based on features (e.g., video clips) that are progressively received from a sensing node (e.g., a camera). Each feature is a temporal sequence of sensory data. The inference error is determined by (i) the timeliness and (ii) the sequence length of the feature, where we use Age of Information (AoI) as a metric for timeliness. While a longer feature can typically provide better inference performance, it often requires more channel resources for sending the feature. To minimize the time-averaged inference error, we study a learning and communication co-design problem that jointly optimizes feature length selection and transmission scheduling. When there is a single sensor-predictor pair and a single channel, we develop low-complexity optimal co-designs for both the cases of time-invariant and time-variant feature length. When there are multiple sensor-predictor pairs and multiple channels, the co-design problem becomes a restless multi-arm multi-action bandit problem that is PSPACE-hard. For this setting, we design a low-complexity algorithm to solve the problem. Trace-driven evaluations demonstrate the potential of these co-designs to reduce inference error by up to 10000 times.

Figures (8)

• Figure 1: A remote inference system, where $X_{t-b}^l:= (V_{t-b}, V_{t-b-1},\ldots, V_{t-b-l+1})$ is a feature with sequence length $l$.
• Figure 2: Performance of wireless channel state information prediction: (a) Inference error Vs. Feature length and (b) Inference error Vs. AoI.
• Figure 3: Performance of actuator state prediction in the OpenAI CartPole-v1 task under mechanical response delay: (a) Inference error Vs. Feature length and (b) Inference error Vs. AoI.
• Figure 4: A multiple source-predictor pairs and multiple channel remote inference system.
• Figure 5: Single Source Case: Time-averaged inference error vs. the scale parameter $\alpha$ in transmission time $T_i(l)=\lceil \alpha l \rceil$ for all $i$.

Theorems & Definitions (8)

• Lemma 1
• proof
• Definition 1: $\epsilon$-Markov Chain ShisherMobihocShisher2023Timely
• Lemma 2
• Theorem 1
• Theorem 2
• Theorem 3
• Lemma 3