人工智能可解释性评估研究综述

Research Progress of Artificial Intelligence Interpretability Evaluation

关键词:  可解释性评估  人工智能可解释性  主观评估  客观评估  评估方法  神经网络  深度学习

Key words:  Interpretability evaluation  Artificial intelligence interpretability  Subjective evaluation  Objective evaluation  Evaluation method  Neural network  Deep learning

摘要:

近年来，可解释人工智能(XAI)发展迅速，成为当前人工智能领域的研究热点，已出现多种人工智能解释方法。如何量化评估XAI的可解释性以及解释方法的效果，对研究XAI具有重要意义。XAI的可解释性评估涉及主、客观因素，是一个复杂且有挑战性的工作。综述了XAI的可解释性评估方法，首先，介绍了XAI的可解释性及其评估的概念和分类；其次，总结和梳理了一些可解释性的特性；在此基础上，从可解释性评估方法和可解释性评估框架两方面，综述和分析了当前可解释性评估工作；最后，总结了当前人工智能可解释性评估研究的不足，并展望了其未来发展方向。

Abstract:

In recent years, explainable artificial intelligence(XAI) has developed rapidly and is becoming a research hotspot in the field of artificial intelligence. Many artificial intelligence explanation methods have emerged. How to quantitatively evaluate the interpretability of XAI and the effect of explanation methods are of great significance to the study of XAI. The evaluation of XAI's interpretability involves subjective and objective factors, which is a complex and challenging task. This paper reviews the interpretability evaluation methods of XAI. Firstly, the concepts and classifications of XAI's interpretability and its evaluation are introduced. Secondly, we summarize and sort out some characteristics of XAI's interpretability. On this basis, current studies on interpretability evaluation are reviewed and analyzed from both aspects of interpretability evaluation method and interpretable assessment framework. Finally, shortcomings of current studies on XAI's interpretability evaluation are summarized, and the future development directions are prospected.