Zebras and antelopes: category sparsity as the result of the relations between objects and within categories*

In their paper “Recognizing a zebra from its stripes and the stripes from ‘zebra’: the role of verbal labels in selecting category relevant information”, Perry and Lupyan (P&L) argue that sparse categories impose high selective attention demands, requiring one to choose what to attend to, compared to dense categories for which several dimensions can or must be used. Furthermore, P&L argue that labels are more useful for sparse categories because they “tune” perception towards the discriminative properties of objects. P&L show that categories for which there is substantial agreement by participants on what the common feature of that category is, have lower selective attention demands and benefit more from the inclusion of a label. In this commentary we focus on what constitutes a sparse category. We will attempt to make the case for the importance of considering both how many discriminative features as well as how many common but not discriminative features a category has in order to evaluate category sparsity. We propose that for a complete understanding of the selective attention processes and labelling effects on category learning one must take the categorisation space, rather than each isolated category, into account. A sparse category is a category for which there are only one or few discriminative features and various irrelevant ones. Increasing the variability in the features among items of the same category (the number of features irrelevant for categorisation) increases sparsity, whereas increasing the number of features varying between the categories (the number of features relevant for categorisation) reduces it (Kloos & Sloutsky, 2008). To identify the sparsity of a category, P&aL asked participants to list the common features to a set of 10 items of the same category, for example 10 pictures of zebras. When asked to list the features of a set of zebras, people are likely to list stripes. Stripes could be a good discriminative feature of zebras (not many animals have stripes), and also a highly characteristic property of that category (all zebras have stripes). However, this might not always be the case, depending on the properties of the contrasting categories and the variability among items of the same category. We believe this is an important consideration because P&L’s measure of sparsity takes into account each category in isolation (thus it is not clear how many discriminating features the categories possess), and does not take into account how many features vary among items of the same category. To illustrate the importance of considering the relation between categories for their sparsity, take the following example using different contrasting categories for the category zebra. One would agree that stripes are a good descriptor of zebras (Figure 1(a)), constituting a sufficient and necessary feature to describe this category – therefore zebra would be considered a sparse category. This would also be the case when one thinks of horses (Figure 1(b)) as a contrast group for the zebra category. However, when contrasting zebras with antelopes (Figure 1(c)) or okapis (Figure 1(d)) stripes are probably not the only good descriptor. In this case, zebra would constitute a denser category. Although people could still frequently list stripes as a commonality, they would also list a series of other features (potentially colour, horns and size), as well as one or more discriminating features. People might still agree that stripes is a good descriptor of zebra but it would no longer be the only or most frequently listed feature. Importantly, the category zebra did not change, it was the space of categories that changed,