Adaptive Splitting Algorithm for Neural Network Modelling in AMP

Artificial Intelligence algorithms usually consume a lot of energy and computing resources during training, so data is usually sent to high-performance devices to complete model training. However, such a computing mode cannot make good use of computing resources in the network and is not conducive to the privacy protection of user information. In recent years, some studies have used end devices to participate in model training, among which federated learning is a typical representative. Due to the heterogeneous computing resources of the end devices, how to split the neural network model to achieve an effective balance between training time and energy consumption is an important issue. In this paper, we propose a novel model splitting method that includes the steps of collecting data, correcting the data, and determining the splitting layers to complete the splitting learning. We consider four factors (training time, energy consumption, CPU utilisation, and memory utilisation) to jointly decide the splitting layer and propose a scheme for adaptively revising the model. We use the VGG16 model for evaluation. The results show that the efficiency of our method is improved by 64.5% compared to the recently proposed ARES scheme. At the same time, our proposed method does not affect the convergence and accuracy of the model.