Knowledge Commons of Institute of Automation,CAS
ProxyBNN: Learning Binarized Neural Networks via Proxy Matrices | |
He, Xiangyu1,2; Mo, Zitao1,2; Cheng, Ke1,2; Xu, Weixiang1,2; Hu, Qinghao1,2; Wang, Peisong1,2; Liu, Qingshan3; Cheng, Jian1,2 | |
2020 | |
会议名称 | European Conference on Computer Vision |
会议日期 | 2020 |
会议地点 | Online |
摘要 | Training Binarized Neural Networks (BNNs) is challenging due to the discreteness. In order to efficiently optimize BNNs through backward propagations, real-valued auxiliary variables are commonly used to accumulate gradient updates. Those auxiliary variables are then directly quantized to binary weights in the forward pass, which brings about large quantization errors. In this paper, by introducing an appropriate proxy matrix, we reduce the weights quantization error while circumventing explicit binary regularizations on the full-precision auxiliary variables. Specifically, we regard pre-binarization weights as a linear combination of the basis vectors. The matrix composed of basis vectors is referred to as the proxy matrix, and auxiliary variables serve as the coefficients of this linear combination. We are the first to empirically identify and study the effectiveness of learning both basis and coefficients to construct the pre-binarization weights. This new proxy learning contributes to new leading performances on benchmark datasets. |
收录类别 | EI |
七大方向——子方向分类 | AI芯片与智能计算 |
文献类型 | 会议论文 |
条目标识符 | http://ir.ia.ac.cn/handle/173211/40619 |
专题 | 复杂系统认知与决策实验室_高效智能计算与学习 紫东太初大模型研究中心_图像与视频分析 |
通讯作者 | Cheng, Jian |
作者单位 | 1.Institute of Automation, Chinese Academy of Sciences 2.University of Chinese Academy of Sciences 3.Nanjing University of Information Science and Technology |
第一作者单位 | 中国科学院自动化研究所 |
通讯作者单位 | 中国科学院自动化研究所 |
推荐引用方式 GB/T 7714 | He, Xiangyu,Mo, Zitao,Cheng, Ke,et al. ProxyBNN: Learning Binarized Neural Networks via Proxy Matrices[C],2020. |
条目包含的文件 | 下载所有文件 | |||||
文件名称/大小 | 文献类型 | 版本类型 | 开放类型 | 使用许可 | ||
He2020_Chapter_Proxy(2540KB) | 会议论文 | 开放获取 | CC BY-NC-SA | 浏览 下载 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论