Server-edge Device Based Efficient Deep Learning Network Retraining Framework

Despite the excellent performance of deep learning, there are two constraints when it is applied to industrial applications: the size of the network and the ability to adapt to unseen data. Recently, attempts have been made to mitigate these issues using a server-edge–device-based environment. In this study, we proposed a server-edge–device-based efficient retraining framework. First, the difficulty score of the unseen data, which was not used in the current training and had no annotation information, was estimated using a deep learning model in the edge device. Thereafter, the filtered unseen data based on the difficulty score were predicted using the labeling model in the server. Finally, retraining and network compression were performed to enhance the performance of the edge model for unseen data. Extensive experiments show that our framework can improve the recognition performance for unseen data and reduce the retraining time by less than three times compared to using the entire dataset, with a performance gap of less than 1%. Using the proposed framework, systematic efficiency can also be achieved by minimizing the data transfer between the server and edge devices.