Abstract:Split learning is an emerging distributed machine learning paradigm that exhibits great advantages in privacy protection and training efficiency. Most previous studies pay attention to client-server based split learning on Convolutional Neural networks (CNNs), and little effort has been spent on investigating the Recurrent Neural Networks (RNNs) oriented split learning. To fill in this gap, we propose a multi-hop split learning algorithm for RNN in mobile edge computing, namely M-SLRNN. The core idea is to efficiently orchestrate resource-constrained mobile devices to collaboratively train a large-size model. We formulate the optimization problem with the objective of minimizing the total training time and theoretically analyze its hardness. Then, we propose an efficient algorithm that conditionally has an approximation ratio of 2. Extensive experiments on both image and text classification tasks demonstrate the effectiveness and the efficiency of the proposed algorithm.

An Efficient Split Learning Framework for Recurrent Neural Network in Mobile Edge Environment.