Perceptual categorisation, Bayesian inference and psychological similarity
Poth, Nina Laura
At the heart of this thesis is the following question: why do we categorise two objects (e.g., an apple and a banana) as instances of the same concept (e.g., the concept fruit) despite their perceptual differences? This is the problem of perceptual categorisation. One way of dealing with this problem is to appeal to the notion of psychological similarity: the apple and the banana belong to the same concept because they look similar. However, there is no scientific agreement on what entity or mechanism the notion of psychological similarity refers to and how this notion explains our ability to categorise both objects as fruit. A promising alternative approach to the problem is Bayesian modelling, whereby perceptual categorisation is analysed as a generalisation and concept-learning task: when categorising the apple and the banana as fruit, we compute the conditional probability that the banana is an instance of the concept fruit, given the background knowledge that the apple is an instance of this concept. This thesis argues for a combination of a Bayesian and a similarity-based approach to perceptual categorisation. I argue that a Bayesian model of concept learning by Tenenbaum and Griffiths (2001) can help us to comprehend a variety of behaviours associated with perceptual categorisation. These were difficult to understand in light of two previous competing theories of psychological similarity—Shepard’s (1987) geometric and Tversky’s (1977) feature-matching theories. One of the behaviours that the Bayesian model can help us comprehend is the tendency to, for example, seek out mushrooms that look similar to edible ones and avoid those that look different from edible ones. The Bayesian model can help us understand why this tendency becomes stronger or weaker depending on how similar or different the mushrooms are. Another of these behaviours is a ‘directionality effect’: we are sometimes more likely to judge Tel Aviv to be similar to New York than vice versa. I argue that the Bayesian approach predicts, systematises and summarises the data on both types of behaviours, whereby it becomes a useful tool to understand perceptual categorisation as a unified phenomenon. The second argument is that the advocated Bayesian approach implicitly relies on a theory of psychological similarity when characterising the hypotheses in the Bayesian inference of perceptual categories. The role of such a similaritybased theory is to explain how a concept such as fruit should be represented in a Bayesian model and how this concept’s representational content is active in producing the subjective probabilities that are associated with hypotheses in a Bayesian inference task.