Split Computing and Early Exiting for Deep Learning Applications: Survey and Research Challenges

Mobile devices such as smartphones and autonomous vehicles increasingly rely on deep neural networks (DNNs) to execute complex inference tasks image classification speech recognition, among others. However, continuously executing the entire DNN mobile can quickly deplete their battery. Although task offloading cloud/edge servers may decrease device’s computational burden, erratic patterns in channel quality, network, edge server load lead a significant delay execution. Recently, approaches based split computing (SC) have been proposed, where is into head tail model, executed respectively device server. Ultimately, this reduce bandwidth usage well energy consumption. Another approach, called early exiting (EE), trains models embed multiple “exits” earlier architecture, each providing higher target accuracy. Therefore, tradeoff between accuracy be tuned according current conditions or application demands. In article, we provide comprehensive survey of state art SC EE strategies by presenting comparison most relevant approaches. We conclude article set compelling research challenges.