Compression of Activation Signals from Split Deep Neural Network

Abstract

The use of artificial neural networks for the purpose of image classification, together with the advancement in computational capabilities of edge devices, plays an important role in the new emerging 5G use case scenarios. However, one of the main challenges of applications involving the use of these networks in edge devices is still the limitation of computational resources. An alternative for saving resources and promoting privacy are the split learning (or split inference) techniques, in which a deep neural network is cut into two parts and executed in distinct devices. Most of these techniques rely on sending the activation signals (output of the cut layer) through the communication channel. This work proposes a new compression algorithm for decreasing the bit rate required for the transmission of the activation signals (or simply “scores”). The presented results demonstrate that the transmission rate can be decreased without hurting the neural network accuracy.