机器学习21：可解释机器学习（Explainable Machine Learning）（上）

摘要

        本周以可解释机器学习为核心议题，系统阐述其在现代人工智能应用中的关键作用。内容重点解析可解释性的重要性——如避免“聪明的汉斯”式表面智能、满足法律合规与公平性要求，并深入探讨模型可解释性与性能强大性之间的权衡关系（如线性模型可解释性强但限制大、深度模型性能优但解释性差）。进一步，将可解释机器学习分类为局部解释与全局解释，并分别介绍其方法与案例，包括显著图、平滑梯度、可视化与探针等技术，揭示模型决策依据，为模型修正与优化提供依据。

Abstract

        This week focuses on Explainable Machine Learning as a core topic, systematically elaborating its critical role in modern AI applications. The content highlights the importance of explainability—such as avoiding "Clever Hans"-style superficial intelligence and meeting legal compliance and fairness requirements—while delving into the trade-off between model interpretability and performance power (e.g., linear models are interpretable but limited, deep models are powerful but less interpretable). Furthermore, explainable machine learning is categorized into local and global explanations, with methods and cases introduced respectively, including techniques like saliency maps, SmoothGrad, visualization, and probing, to reveal model decision rationale and support model correction and optimization.

一．可解释机器学习重要性

        前面学习了许多模型如给出一张图片可以告诉我们图片内容的影像辨识的模型，但是我们并不能止步于此，接下来要让机器给出得到答案的理由。

        可解释机器学习是一个重要议题是因为机器即使能够得到正确答案但是并不代表其就非常聪明。就