Research Area: Machine Learning
Deep neural networks(DNNs) have achieved great success in many real-world applications, but they also had some drawbacks such as considerable storage requirement, large computational power consumption and delay for training and inference, making it impracticable to deploy state-of-the-art models into embedded systems and portable devices. Thus, the demand of compressing DNNs has been taken into consideration. In this paper, we focus on quantized neural networks, which is one scheme of compressing DNNs. At first we introduce some baseline works in quantized neural networks and then give a review on optimization ways used in quantizing neural networks. In our perspective, these methods fall into two categories: minimizing quantization error and minimizing loss function. Specialized introduction for each category follows after baseline works. We also make some comments to each category and some methods. Finally, we discuss on some possible directions of this area and make a conclusion.
Keywords:
Author(s) Name: Yuhan Lin; Minglong Lei; Lingfeng Niu
Journal name:
Conferrence name: International Conference on Data Mining Workshops (ICDMW)
Publisher name: IEEE
DOI: 10.1109/ICDMW.2019.00063
Volume Information:
