基于深度学习和多源数据的自动驾驶技术风险识别

Risk identification of autonomous driving technologies based on deep learning and multi-source data

新兴技术发展带来创新和市场机会同时,也给科技治理带来了新挑战。科林格里奇困境理论认为技术的社会治理之所以存在障碍是因为当该技术负外部性涌现时,已经与人类社会深度绑定。因此,如何在新兴技术发展的早期或前期及时发现或预见其潜在风险,对于相关科技政策制定和完善具有积极意义。针对新兴技术发展早期风险信号较为稀疏和难以识别是科学问题,本文从研究者知识、企业认知和个体感知三个维度出发,综合运用论文、专利、企业年报和社交媒体数据,构建了一个量化的新兴技术风险识别框架,并以自动驾驶技术为例进行了实证检验。理论和实证研究发现:基于深度学习和多源数据的分析框架可以进一步挖掘新兴技术的潜在风险信号,为科技风险治理提供更为丰富的决策参考;其中,Siamese-BERT孪生网络模型在多种深度学习方法中对新兴技术风险信号的挖掘效果较好,实证分析则发现自动驾驶技术潜在风险涉及诸多不同领域,需引起相关政策和监管部门的关注。本文提出了新兴技术风险分析的知识基础理论,以及基于深度学习的量化分析框架,从知识管理和文本语义挖掘视角进一步丰富了当前有关新兴技术管理的理论和方法体系,针对如何在新兴技术演进的早期阶段进行潜在风险评估,提供了新的观察视角和量化工具。

The development of emerging technologies not only brings innovation and market opportunities,but also brings new challenges for science and technology governance.The theory of Collingridge dilemma suggests that the social governance of technology is hindered because when negative externalities of a specific technology emerge,it has become deeply entangled with human society.Therefore,how to identify or foresee the potential risks of emerging technologies in the early or initial stages is of great significance for the formulation and improvement of relevant science and technology policies.Identifying the early risks of emerging technologies is a scientific problem due to the sparse and difficult-to-identify signals.Based on the proposed three dimensions of researcher knowledge,enterprise cognition and individual perception,this paper used academic papers,patents,enterprise annual reports,and social media data to construct a quantitative framework for identifying emerging technology risks.An empirical test was conducted by using the example of autonomous driving technologies.The theoretical and empirical analyses showed that the integrated analysis framework based on deep learning and multi-source data can further mine potential risk signals of emerging technologies,thus providing more diverse decision-making references for science and technology risk governance.Among various deep learning methods,the Siamese-BERT twin network model has better risk signal mining effects for emerging technologies.Regarding autonomous driving technology,the empirical analysis showed that potential risks involving many different fields require attention from relevant policy and supervision departments.This paper has presented a theoretical and quantitative analysis framework based on deep learning for emerging technology risk analysis,which has further enriched the current theoretical and methodological system of emerging technology management from the perspectives of knowledge management and text semantic mining,and it will provide a new perspective and tools for potential risk assessment in the early stages of emerging technology evolution.