Improving Edge Collaborative Inference Systems with Mixed Criterion NN

In this research paper, the authors explore the potential benefits of using edge computing to improve the performance of deep neural networks (DNNs) in mobile devices. They propose two new transfer decision (TD) mechanisms that can be used to determine when to transmit data from a mobile device to an edge server for further processing. The first mechanism, called confidence threshold, is based on the confidence level of the DNN’s predictions, while the second mechanism, called entropy threshold, is based on the entropy of the softmax output of the classifier.
The authors evaluate these two mechanisms using a dataset of images and show that they can achieve significant communication savings without compromising accuracy. They also compare their approach to other state-of-the-art TD mechanisms and demonstrate its superiority in terms of communication efficiency.
To explain their findings, the authors use analogies such as "a complex ML algorithm is split into two parts" and "DNNs trained for a specific task learn to only preserve the information essential for that task." They also highlight the advantages of executing an initial part of the DNN on the edge device, which can reduce communication overhead and improve accuracy.
Overall, the authors demonstrate that TD mechanisms can play an important role in improving the performance of DNNs in mobile devices by balancing accuracy and communication efficiency. Their proposed methods have the potential to be applied in a wide range of applications, including image classification, object detection, and natural language processing.