Within-Category Feature Correlations and the Curse of Dimensionality

Abstract

Within the domain of category learning, the curse of dimensionality states that as categories acquire more features, the size of the feature space and thus the number of examples necessary to adequately learn the category grows rapidly. As a result, category learning should be a highly difficult task. However, people learn to classify categories with ease. The primary aim of the current study was to address how people overcome this problem by determining if they were attuned to information about which features were correlated with each other, and whether or not they used this information as an indication that those features were relevant for categorisation. In theory, this would then allow for the detection of natural category family resemblance structure in highly dimensional environments, and subsequently, allow human learners to overcome the curse of dimensionality. In addition to this primary aim, and under the assumption that people were attuned to this correlational structure, a number of secondary hypotheses were also proposed. These hypotheses assessed both feature and correlation learning, evaluated relative category learning improvement as a function of dimensionality between conditions, as well as compared human performance to a Naïve Bayes model that was inherently incapable of detecting any correlational structure. The results of the current study suggest that people do not utilise within-category feature correlational structure as a heuristic for category predictive feature detection. However, these findings were contingent on a number of methodological shortcomings present in the current study. Directions for future research are proposed that provide clear methodological changes in category structure which may mitigate the lack of support found for the proposed hypotheses.