CTO(首席技术官)英文Chief Technology Officer,即企业内负责技术的最高负责人。

VIP内容

题目:

图神经网络与认知推理

摘要:

图神经网络将深度学习方法延伸到非欧几里得的图数据上,大大提高了图数据应用的精度。我将首先从算法角度分析当下经典的图表示学习算法(DeepWalk、LINE、node2vec等)的本质关系,并提出统一算法框架NetMF和大规模版本NetSMF,并在稀疏图理论的基础上提出高效快速学习算法ProNE,ProNE在精度不降低的情况下比传统学习算法快10-400倍的加速比。接着,我会简单回顾一下图卷积网络(GCN)并探讨如何提高GCN在图数据上的表示学习能力。我们研究发现几个巧妙、简单方法就可以有效的提高GCN的表示能力,该方法可以等价表示为图注意力网络(GAT)。该方法在包括阿里巴巴等多个超大规模数据集上得到应用验证。最后我将探讨在图神经网络基础上的认知推理模型CognitiveGraph (CogGraph)。CogGraph基于认知科学中的双通道认知理论,其中通道1负责直觉认知,通道二负责推理认知。CogGraph可以广泛应用于多个图数据上的任务,包括基于推理的问答、知识图谱补齐等。

个人简介:

张鹏,北京智谱华章科技有限公司CTO,清华大学2018创新领军工程博士,毕业于清华大学计算机科学与技术系知识工程研究室,研究领域包括文本数据挖掘和语义分析、知识图谱构建和应用等。长期致力于将语义信息挖掘和知识图谱技术应用于各种行业应用,在语义大数据分析、智能问答、辅助决策等应用领域拥有多年实践经验。

成为VIP会员查看完整内容
0
53

最新论文

Organizations are adopting data analytics and Business Intelligence (BI) tools to gain insights from the past data, forecast future events, and to get timely and reliable information for decision making. While the tools are becoming mature, affordable, and more comfortable to use, it is also essential to understand whether the contemporary managers and leaders are ready for Data-Driven Decision Making (DDDM). We explore the extent the Decision Makers (DMs) utilize data and tools, as well as their ability to interpret various forms of outputs from tools and to apply those insights to gain competitive advantage. Our methodology was based on a qualitative survey, where we interviewed 12 DMs of six commercial banks in Sri Lanka at the branch, regional, and CTO, CIO, and Head of IT levels. We identified that on many occasions, DMs' intuition overrules the DDDM due to uncertainty, lack of trust, knowledge, and risk-taking. Moreover, it was identified that the quality of visualizations has a significant impact on the use of intuition by overruling DDDM. We further provide a set of recommendations on the adoption of BI tools and how to overcome the struggles faced while performing DDDM.

0
0
下载
预览
Top