The epistemology of machine learning: From bias to knowledge

当代人工智能方法的激增要求对所涉及的认识论问题进行哲学分析。这种分析的一个特别紧迫的主题是机器学习中的偏见。在使用机器学习方法的各个阶段，算法偏差都会出现。值得注意的是，在当前的工作中相对被忽视的是实际学习发生的阶段：机器学习算法从训练数据中进行泛化。对不可避免的归纳偏差的研究是机器学习理论的领域，但是这种数学方法还没有提供归纳偏差概念的清晰概念图。缺少的是对归纳偏差的统一认识论解释，这种认识论将学习理论中的各种技术分析纳入其中。根据认知成功的条件来解释归纳偏差，是一个自然的起点，可以更普遍地解释我们如何通过机器学习方法获得知识。这种认识论的现有概要倾向于挑出机器学习方法使用的经验性质，而形式哲学中的许多工作在关注学习代理的理想合理性方面采取了另一个极端。我们仍然缺少的是一种对机器学习的实践和理论都公正的描述。这个项目的目的是填补这两个空白：发展归纳偏见概念的哲学解释，并将其纳入新的机器学习认识论。这项工作通过两个广泛的案例研究得到了加强，这些案例研究将这一解释应用于围绕机器学习方法的当代核心辩论。因此，该项目由四个工作包组成。首先，我们激发和发展归纳偏见的解释。主要思想是根据方法成功学习的条件和性质进行一般认识论表征，对于每个特定的方法都可以用机器学习理论进行精确描述。其次，我们利用我们的解释来推进关于解释深度神经网络的偏差和经验成功的辩论。我们进一步评估我们的解释，并确定可能的概念限制。第三，我们使用我们的解释方案来推进关于算法公平性的争论。我们描述了归纳偏差中认知因素和非认知因素之间的关系，以及归纳偏差和算法偏差之间的一般相互作用。最后，我们发展了机器学习方法的实用主义认识论。主要的观点是，对归纳偏见的中介作用的关注自然地与对探究本质的看法保持一致，这是科学哲学中皮尔斯实用主义传统的核心。

The contemporary surge of methods in artificial intelligence calls for philosophical analysis of the epistemological issues involved. A particularly pressing topic for such analysis is bias in machine learning. There are various stages in the use of machine learning methods where algorithmic bias enters. Relatively neglected in current work, remarkably, is the stage where the actual learning takes place: where a machine learning algorithm generalizes from the training data. The study of the inevitable *inductive bias* that arises here is the domain of machine learning theory, but this mathematical approach does not yet provide a clear conceptual picture of the notion of inductive bias. What is missing is a unified epistemological account of inductive bias that subsumes the various technical analyses from learning theory. An account of inductive bias in terms of conditions for epistemic succes is a natural starting point for a more general account of how we gain knowledge through machine learning methods. Existing sketches of such an epistemology tend to single out the empirical nature of the use of machine learning methods, while much work in formal philosophy takes the other extreme in its focus on the ideal rationality of learning agents. What is still missing is an account that does justice both to the practice and the theory of machine learning. The aim of this project is to fill these two gaps: to develop a philosophical explication of the concept of inductive bias, and to incorporate this into a novel epistemology of machine learning. This work is reinforced by two extensive case studies that apply this explication to central contemporary debates around machine learning methods. The project thus consists of four work packages. First, we motivate and develop an explication of inductive bias. The leading idea is a general epistemological characterization in terms of the conditions and nature of a method’s successful learning, that for each specific method can be made precise with machine learning theory. Second, we employ our explication to advance the debate about the explanation of the biases and empirical success of deep neural networks. We further evaluate our explication, and identify possible conceptual limitations. Third, we employ our explication scheme to advance the debate on algorithmic fairness. We delineate the relations between epistemic and nonepistemic factors in inductive bias, and the interactions between inductive bias and algorithmic bias in general. Finally, we develop a pragmatist epistemology of machine learning methods. The leading idea is that a focus on the mediating role of inductive bias naturally aligns with a view on the nature of inquiry that is core to the Peircean pragmatist tradition in the philosophy of science.