Dept. of Psychology, Center for Cognitive Science, Rutgers University, United States.
Dept. of Philosophy, Rutgers University, United States.
Cognition. 2022 Jun;223:105058. doi: 10.1016/j.cognition.2022.105058. Epub 2022 Feb 24.
The rise of Bayesian models of cognition requires that traditional questions in epistemology and metaphysics, such as how models relate to reality and how one observer's models relate to another's, be reframed in probabilistic terms. In this paper we take up these questions beginning from a subjective (Bayesian) conception of probability, in which distinct observers hold potentially different probabilistic models of the world, with no one observer necessarily possessing the "true" one. The key question is what terms in a probabilistic theory mean-that is, what they refer to and what their truth conditions are. We address this question with tools from information theory. We introduce the translation uncertainty, a generalization of the Kullback-Leibler divergence that expresses the discrepancy between two observers' probabilistic models of a common environment. We derive a number of basic information-theoretic relationships among observers, showing for example that the probability that two Bayesian observers will classify the world similarly (called the concordance) depends on the translation uncertainty between their respective models of the world. Our framework suggests a pathway to a semantics for a "probabilistic language of thought."
贝叶斯认知模型的兴起要求认识论和形而上学中的传统问题，如模型与现实的关系以及一个观察者的模型与另一个观察者的模型的关系，用概率术语重新表述。在本文中，我们从主观（贝叶斯）概率的概念出发来处理这些问题，在这个概念中，不同的观察者持有潜在不同的世界概率模型，没有一个观察者必然拥有“真实的”模型。关键问题是概率理论中的术语是什么意思——也就是说，它们指的是什么，它们的真值条件是什么。我们使用信息论的工具来解决这个问题。我们引入了翻译不确定性，它是 Kullback-Leibler 散度的推广，表达了两个观察者对共同环境的概率模型之间的差异。我们推导出了观察者之间的一些基本信息论关系，例如，两个贝叶斯观察者将世界分类相似的概率（称为一致性）取决于他们各自世界模型之间的翻译不确定性。我们的框架为“概率思维语言”的语义学提供了一条途径。
