Split Neural Networks for Mobile Devices

Abstract

In some areas neural networks are becoming a non-alternative way of solving problems. Recognizing images, sounds, classification - these problems require serious processor power and memory for neural network learning and operation. Modern mobile devices have quite good characteristics for executing only the primary layers of deep neural networks, but there are not enough resources for entire networks. Since the training of non-mobile networks for mobile devices takes place separately on external resources, a method was developed for the distributed operation of a neural network with vertical partitioning sets of layers with synchronization of learning data. In some cases, the proposed approach allows to use full-sized deep neural networks on mobile device, where they are needed without overloading the communication channel and device resources.