人工智能风险分析技术研究进展

Research progress on risk analysis for artificial intelligence

目前基于深度学习模型的预测在真实场景中具有不确定性和不可解释性,给人工智能应用的落地带来了不可避免的风险。首先阐述了风险分析的必要性以及其需要具备的3个基本特征:可量化、可解释、可学习。接着,分析了风险分析的研究现状,并重点介绍了笔者最近提出的一个可量化、可解释和可学习的风险分析技术框架。最后,讨论风险分析的现有以及潜在的应用,并展望其未来的研究方向。

The predictions of the deep learning models are still uncertain and uninterpretable.As a result,their deployments bring unavoidable risk to business decision making.Firstly,the study on risk analysis was motivated,and the three desirable properties of risk analysis techniques were described:quantifiability,interpretability and learnability.Then the existing work on risk analysis was reviewed,and the newly proposed framework to enable quantifiable,interpretable and learnable risk analysis was introduced.Finally,the existing and potential applications of risk analysis,and its future research direction were discussed.