Effects of labelling on object perception and categorisation in infants

Capelier-Mourguy, Arthur and Westermann, Gert and Twomey, Katherine (2019) Effects of labelling on object perception and categorisation in infants. PhD thesis, Lancaster University.

Text (2019capeliermourguyphd)
2019capeliermourguyphd.pdf - Published Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (7MB)

Abstract

How do labels impact object perception and enhance categorisation? This question has been the focus of substantial theoretical debate, particularly in the developmental literature, with conflicting results. Specifically, whether labels for objects act as additional perceptual features or instead as referential pointers to category concepts has been the subject of intense debate. In this thesis, we attempted to shed a new light on this question, combining empirical results on both infants and adults, and neurocomputational models. First, we developed a dual-memory neurocomputational model of long-term learning inspired by Westermann and Mareschal's (2014) model, to test predictions of the two mains theories on labelling and categorisation on existing infant data, and to generate predictions for a follow-up study. Our modelling work suggested that for the empirical designs considered and age groups tested, labels were processed as object features, as opposed to having a more referential role. We then focused on explicitly testing potential attentional effects of auditory labels during categorisation in an empirical study. More precisely, we studied the interaction between feature salience, feature diagnosticity, and auditory labels, in a categorisation task. Surprisingly, we found that 15-month-old infants and adults could learn labelled categories in which the salient feature (head of line-drawn novel animals) was non-diagnostic of category membership, but the non-salient feature (tail) was, without adopting a different pattern of looking compared to participants in a control group. Although our data did not provide clear evidence for a true null effect, this finding was once again more compatible with the theory that labels act as features, not referents. This finding also led us to reconsider the use of eye movements and looking times as a proxy for learning, as it seemed that participants could learn more without looking more. Given our empirical results on salience and diagnosticity of features, and given the methodological differences in the handling of feature salience and diagnosticity in the categorisation literature, we developed a simple auto-encoder model to further study the impact of salience differences between features in the context of a categorisation task, with or without a label. Our simulations suggested that bigger disparities in salience between different features of an object can result in differences in terms of learning speed and compactness of categories in internal representations, hinting that future empirical studies should consider feature salience in their design. Overall then, this thesis provides some evidence in favour of the labels-as-features theory through the use of empirical eye-tracking data on infants and adults, and neurocomputational modelling. This thesis further asks new questions on the importance of feature salience in categorisation tasks, and the interpretation of eye movement and looking time data in general.