Hardware-based real-time deep neural network lossless weights compression

Abstract

Deep Neural Networks (DNN) are widely applied to many mobile applications demanding real-time implementation and large memory space. Therefore, it presents a new challenge for low-power and efficient implementation of a diversity of applications, such as speech recognition and image classification, for embedded edge devices. This work presents a hardware-based DNN compression approach to address the limited memory resources in edge devices. We propose a new entropy-based compression algorithm for encoding DNN weights, as well as a real-time decoding method and efficient dedicated hardware implementation. The proposed approach enables a significant reduction of the required DNN weights memory (approximately 70% and 63% for AlexNet and VGG19, respectively), while allowing the decoding of one weight per clock cycle. Results show a high compression ratio compared to well-known lossless compression algorithms. The proposed hardware decoder enables an efficient implementation of large DNN networks in low-power edge devices with limited memory resources.