One-shot Learning of Realistic Categories
Zusammenfassung der Projektergebnisse
The aim of the proiect was to get a better understanding of one-shot category learning. How is it possible that humans can learn new categories from very few examples? Two scenarios were identified that seem well-suited to study this question. The first scenario assumes that there are systems of concepts and that categories interrelate. For example, many categories are organized into a hierarchical ontology. Learning about the superordinate category by learning several basic-level categories will help learning new basic-level categories. The second scenario assumes that objects in a category are highly structured. For example, a plant that is the result of a growth process carries information about the growth process in its structured appearance. The two scenarios can interact, too. Objects can be made from parts and sub-parts that form categories themselves and parts can be re-used in other categories (think of wheels that are parts of cars and bikes). Several sets of artificial stimuli were developed for demonstration purposes but no experiments were run yet. Recurring problems in the development of suitable stimulus material were: (a) There were low-level features that were confounded with the structure that was supposed to be learned, and (b) the more structured the stimuli were, the less clear it was how participants would perceive and represent the stimuli. Hence, finding a formal representation for interestingly structured visual stimuli was identified as a crucial problem to be overcome before theoretical progress could be made. This representation has to be able to represent partwhole structures, coordinate features between parts that can be far apart, and represent symmetry and repetition. This representation would also have to support learning of categories which is more difficult the richer the underlying representations are. While the requirements are now clear we have not found an adequate representation yet. Furthermore, visually presented stimuli with interesting structure always elicited processes of perceptual organization that are not understood very well. Hence, an effort was made to model perceptual organization for one class of stimuli that seemed relevant for category learning but also well-studied by perception scientist (moving light displays). This work is still on-going.
