Neural network quantization algorithms based on weights and activation values
Yubo Li, Runbo Wang, Huaping Wu · 2022
In recent years, neural networks have demonstrated attractive performance in a variety of computer vision tasks. This has led to the deployment of CNNs into real-world applications. However, most existing CNNs require a large amount of storage space and computational resources, which hinders their deployment on mobile devices. In order to solve this problem, the existing research shows that quantization and pruning can greatly reduce the memory occupied by the neural network in the device without reducing the accuracy of image recognition, which plays a great role in solving the deployment problem of CNNs. Therefore, this paper classifies and analyzes the quantitative algorithm model of neural network according to the parameters that need to be quantified, summarizes the common defects of the current quantitative algorithm in terms of accuracy, and looks forward to its future development trend.