Unsupervised categorization is the task of classifying novel stimuli without external feedback or guidance, and is important for every day decisions such as deciding whether emails fall into 'interesting' or 'junk' categories (Gureckis & Love, 2003). This thesis aimed to further investigate this currently under-researched phenomenon by comparing it to the standard, supervised categorization; developing a set of principles upon which people can rely when completing an unsupervised categorization task; and determining how people choose to categorize when all restraints are removed. The first section of the thesis found that when compared to supervised categorization, performance in an unsupervised task was found to be equivalent in terms of trial-by-trial acquisition of strategy. In addition, the post-categorizing effects on perception found in previous literature on supervised categorization were moderately replicated using an unsupervised task. The second part of the thesis examined the putative principles of unidimensionality, balance, frequency and gap in continuity by manipulating stimulus sets, and all were found to have some merit in predicting how participants will choose to categorize in an unsupervised categorization task. Finally, the ultimate experiment within the thesis relaxed all constraints and, contrary to the previous experiments, allowed participants to create as many categories as they wished. Results indicated that when participants chose to create few categories, their learning was significantly better than when large numbers of categories were created. Overall, the results of this experimental series indicate that unsupervised categorization is equivalent to supervised categorization in a number of ways, and that it can also be predicted using the proposed principles of unidimensionality, balance, frequency and gap in continuity; however this is limited when all constraints are removed.
|Qualification||Doctor of Philosophy|
|Publication status||Unpublished - 2008|