Answers are grounded in this paper and cite their sources.
In device–edge split inference, a device sends intermediate features of a deep network to an edge server over a link whose capacity is limited and varies. We propose the Resource-Aware Asymmetric Semantic Codec (RAASC), which adapts the payload by transmitting only a prefix of its latent channels. An encoder with learned channel mixing prepares the latent for truncation, and a decoder conditioned on the retained channel fraction and the supplied bit-error rate (BER) aggregates spatial context from the received prefix through learned global summaries. Multi-budget task-aware training over prefix lengths from four to 32 channels lets one parameter set per task serve different budgets. On CIFAR-100 classification and COCO2017 detection, we compare RAASC with SemanticNN, an error-resilient offloading codec, adapted to the same protocol and given the budget as input. Averaged over five budgets and five BERs, RAASC improves Top-1 accuracy by 1.60 percentage points and mean average precision by 0.88 points, with gains in all three training seeds at every condition. Its 24-channel prefix also outperforms the reference’s 32-channel prefix at every evaluated BER with 25% fewer feature-index bits. This shorter option needs no more total processing time on a single CPU thread; on a desktop GPU at batch one, it is at most 0.42 ms slower, which the saved transmission time offsets on links below about 3.7 and 27 Mbit/s, respectively.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: