LSTM-Enabled Scalable Semantic Retransmission for Device-edge Co-inference Systems

In this paper, we investigate a semantic retransmission policy for a device-edge co-inference (DEC) system, where an edge device (ED) extracts semantic features from raw image data and transmits them to an edge server (ES) via adaptive retransmissions to optimize AI-driven classification tasks. In particular, we propose a scalable long short-term memory (LSTM)based classifier that dynamically aggregates sequential semantic features across transmissions, enabling adaptive adjustments to the maximum retransmission count (e.g., increasing retransmissions in response to stricter accuracy thresholds or deteriorating channel conditions) without model retraining. Building on this framework, we design a deep reinforcement learning (DRL)-based optimal-stopping automatic repeat request (ARQ) policy to jointly optimize classification accuracy and end-to-end transmission delay. Experimental results demonstrate that the LSTM classifier achieves near-optimal accuracy (within $1.3 \%$ of the empirical upper bound), even when retransmissions exceed its training range. Notably, the DRL-based ARQ scheme reduces average latency by $49.7 \%$ compared to conventional bit-level ARQ, with a limited $4.6 \%$ accuracy degradation.