Inductive biases and the geometry of concept learning

Language learning is generally viewed as an inductive process that relies on biases to operate efficiently. How these biases emerge and stabilize during the learning process remains a question that continues to occupy psychologists and philosophers. In this article, we address this issue from the perspective of the theory of conceptual spaces. The main idea is that an inductive bias can be understood as a domain-specific deformation of the learner's conceptual space. We propose a novel model of bias formation wherein learners infer, from labeled observations, which dimensions are most diagnostic within a given naming context. This is achieved by calculating the ratio of between-category variation to within-category variation for each dimension. Furthermore, we explore the broader implications of this model and its relationship to other cognitive phenomena, such as essentialism.