ET-SCS: Error-Tolerant Split Computing System for Internet of Things

Split computing reduces the inference latency of an artificial intelligence (AI) model by offloading part of the model to an edge cloud near an artificial intelligence of things (AIoT) device. However, conventional AI models are not trained for split computing, and transmitting intermediate features over wireless links without error recovery causes accuracy degradation. In this article, we propose an error-tolerant split computing system (ET-SCS), where the edge cloud trains the model by considering transmission errors in intermediate data. ET-SCS also introduces dynamic filtering that automatically adjusts filter coefficients for robustness. Consequently, AIoT devices can select the split point to minimize inference latency or energy consumption without concern about performance loss. Evaluation results show that ET-SCS reduces energy consumption by up to 39% and inference latency by up to 64% compared with the nonsplit model, without significant accuracy degradation.