In nature an animal seldom encounters physically identical stimuli twice, and hence needs rules for deciding whether observed differences are important. In some cases generalization may be innate (and/or low-level), but often it is beneficial to learn about how stimuli vary. We train young poultry chicks about colours and colour variation associated with food, and then test how they respond to and learn about novel colours. A key feature is that chicks are capable of discriminating very small colour differences, and trained on a single example generalise little (about 1nm on the monochromatic spectrum at c. 580nm). This excellent colour vision means that chicks ‘risk’ making inappropriately fine discriminations when foraging for food, and allows us to study how they use information about how object colours vary they are familiar with multiple examples. Our observations suggest that in their initial response to a novel stimulus the chicks make efficient use of the knowledge about stimulus variation acquired during training. The chicks’ then subsequent learning about the novel colours is much greater for the novel stimuli than for the established training colours, which is consistent with Bayesian models in that the rate of learning is related to the uncertainty about the stimulus. Thus a stimulus can simultaneously be similar to familiar stimuli - in that it is predicted to have similar properties -, and at the same time as be recognised as novel, so that the prediction is labile. I will discuss how these observations of chick behaviour can be related to models of generalization that have been applied to animal behaviour, where it seems that ‘similarity’ and ‘novelty’ are mutually exclusive properties.
BACK
