从统计学与心理学的视角看可解释性人工智能

Explainable Artificial Intelligence from Perspective of Statistics and Psychology

人工智能给出结果的可解释性是人工神经网络用户在实际应用中十分关注的一个问题。针对这一问题,对DARPA设立的可解释性人工智能项目的主要研究内容进行了概要性解读,首先从统计学的视角对人工神经网络的任意函数逼近与统计回归的本质进行了分析,指出了人工神经网络所给出结果的概率统计意义。在此基础上从心理学的视角对人与机器之间沟通与说服的形式与过程进行了剖析,展示了心理学在人机交互与人机信任建立过程中所发挥的重要作用。这对于更加理性地看待人工智能的发展以及人工神经网络的应用边界条件具有重要的参考意义。

The explainable result from artificial intelligence is very significantly concerned by users in the application of artificial neural networks.In this paper,the main research contents of the explainable artificial intelligence project set up by DARPA are briefly introduced.Firstly,the essence of arbitrary function approximation and statistical regression for artificial neural network is analyzed from the perspective of statistics.The probabilistic and statistical significance of the results given by artificial neural network is pointed out.The process of communication and persuasion between human and machines is analyzed from the perspective of psychology,and the important role of psychology in human-machine interaction and the trust establishment is demonstrated.It is the significant reference for the artificial intelligence development and artificial neural network application condition in a more rational view.