An important function of category learning is that it allows category knowledge to be applied to new items to determine appropriate action. For instance, my ability to categorize mushrooms as poisonous or non-poisonous has little utility if I cannot apply that knowledge to new mushrooms I come across while foraging.
Extensive work on humans has shown that generalization is easier for categories defined by a single rule about one perceptual dimension (rule-based categories) compared to categories that require integration across perceptual dimensions (information-integration categories).
For example, it may be relatively easy to determine whether a new mushroom is a safe chanterelle or a poisonous Jack O'Lantern because they differ in a single feature (false gills vs. true gills). In contrast, poisonous "little brown mushrooms" can only be distinguished from safe honey mushrooms based on combinations of several features, which may make it more difficult to apply what is learned about a specific little brown mushroom to new examples.
Because the advantage for rule-based categories has been attributed to the fact that the categorization rule can be verbalized and held in memory, it is interesting to ask whether differences in generalization between rule-based and information-integration categories are also evident in nonhuman primates without verbal abilities.
To answer this question, Smith and colleagues (2015, Journal of Experimental Psychology: Animal Learning and Cognition) (PDF, 388KB), taught humans and rhesus macaques to categorize rectangles that varied in size and density of illuminated pixels (roughly translates to brightness). For rule-based category learning, categories were defined by one dimension (e.g., size); for information-integration category learning, categories were defined by a combination of both size and brightness.
During training, stimuli drawn from training regions of the category space were presented one at a time, and subjects had to categorize them as A or B. Feedback was provided, and incorrect responses were penalized with an 8-second timeout period. Following training, humans and rhesus macaques were tested with stimuli drawn from untrained regions of the category space to assess generalization.
Both species successfully learned rule-based and information-integration categories. Importantly, both species also displayed the same pattern of generalization performance. For rule-based categories, generalization performance was as high as performance at the end of training, and model-based analyses of individual subject data indicate that the same decision criteria easily transferred from training to generalization stimuli; in contrast, for information-integration categories, generalization performance was impaired relative to the end of training, and subjects were generally unable to transfer their decision strategy from training to the untrained region of the stimulus space.
These results suggest that, in both species, rule-based category knowledge is independent from the training context. More broadly, this result suggests that nonhuman primates are able to represent dimensional rules in the absence of verbal codes, perhaps via learned attention to the relevant perceptual feature, pointing to a possible origin of the explicit declarative system underlying human rule-based category learning.
Citation:
Smith, J. D., Zakrzewski, A. C., Johnston, J. J. R., Roeder, J. L., Boomer, J., Ashby, F. G., & Church, B. A. (2015). Generalization of category knowledge and dimensional categorization in humans (Homo sapiens) and nonhuman primates (Macaca mulatta). Journal of Experimental Psychology: Animal Learning and Cognition, 41(4), 322–335. http://dx.doi.org/10.1037/xan0000071
Note: This article is in the Basic / Experimental Psychology topic area. View more articles in the Basic / Experimental Psychology topic area.