Tensor-Empowered Hardware-Friendly Lightweight Deep Neural Networks for Vehicular Edge Computing

The high memory footprint, high computational overhead and high power consumption of deep neural networks are the main bottlenecks in deploying network models to vehicular edge devices. Furthermore, in the collaborative learning process, the large number of training parameters can cause high communication overhead between the vehicular edge devices and the cloud. The redundant floating-point operations and training parameters are the main “culprits” for these bottlenecks. To solve the above problems, we propose the lightweight tensor linear shift layer and the lightweight tensor convolutional shift layer, and two models based on the lightweight tensor shift layer, LTS- $\alpha $ and LTS- $\beta $ are proposed. Firstly, the weight matrix and weight tensor in the network model are represented in the form of a chain-structured weight tensor kernel; secondly, fewer bits are used to represent the values in the weight tensor kernel; finally, the floating-point multiplication operations in the network model are replaced with inexpensive sign flipping and bitwise shift, as a way to reduce the amount of floating-point computation and memory footprint in the deep neural network model, and to speed up the training and inference process of the model, thereby reducing the power consumption of the model at runtime. To evaluate the superiority of our proposed model, we conduct experiments on several real datasets and compare them with the relevant mainstream method. The experimental results show that both our proposed LTS- $\alpha $ and LTS- $\beta $ can achieve comparable or even higher performance on real tasks than the relevant mainstream method. Moreover, LTS- $\alpha $ and LTS- $\beta $ have a lower memory footprint and computational overhead than relevant mainstream methods, and the training process and inference process of the models consume less energy. Therefore, LTS- $\alpha $ and LTS- $\beta $ are more suitable for deployment to resource-constrained vehicular edge devices to provide more general intelligence services.