An Overview on Lightweight Methods for Deep Neural Network Parameters
In recent years,deep neural networks have achieved great success on a variety of challenging tasks,constantly refreshing people's understanding of artificial intelligence.However,the deep neural network model is difficult to deploy to resource-limited edge computing devices because of its huge param-eters,high computing and storage costs.Therefore,people begin to try to lightweight network design from the perspective of network architecture and parameter number,while ensuring the performance of neural network is acceptable.In this paper,the development history and working principle of deep neural net-works are reviewed briefly from the angle of network parameter lightweight.Secondly,three kinds of cur-rent mainstream parameter lightweight methods are introduced:parameter quantization,tensor decomposi-tion and parameter sharing.Then the advantages and disadvantages of the method are compared and ana-lyzed from the dimensions of thought description,application level and training mode.Finally,the future development direction of neural network lightweight is prospected.
deep neural networksartificial intelligenceedge computing devicesneural networks pa-rameters lightweight
NETL
NSTL
万方数据
