Improving the Accuracy-Latency Trade-off of Edge-Cloud Computation Offloading for Deep Learning Services

Offloading tasks to the edge or the Cloud has the potential to improve accuracy of classification and detection tasks as more powerful hardware and machine learning models can be used. The downside is the added delay introduced for sending the data to the Edge/Cloud. In delay-sensitive applications, it is usually necessary to strike a balance between accuracy and latency. However, the state of the art typically considers offloading all-or-nothing decisions, e.g., process locally or send all available data to the Edge (Cloud). Our goal is to expand the options in the accuracy-latency trade-off by allowing the source to send a fraction of the total data for processing. We evaluate the performance of image classifiers when faced with images that have been purposely reduced in quality in order to reduce traffic costs. Using three common models (SqueezeNet, GoogleNet, ResNet) and two data sets (Caltech101, ImageNet) we show that the Gompertz function provides a good approximation to determine the accuracy of a model given the fraction of the data of the image that is actually conveyed to the model. We formulate the offloading decision process using this new flexibility and show that a better overall accuracy-latency trade-off is attained: 58% traffic reduction, 25% latency reduction, as well as 12% accuracy improvement.