We consider the application of symbolic learning in natural domains. Given a conceptualization of the domain in terms of a particular set of predicates, we consider how knowledge about the domain can constrain learning. In particular, by specifying a set of predicates which are not allowed to vary within a category we restrict the set of possible categories. We suggest that such constraints arise naturally in perception if we distinguish the set of "singular" predicates which describe degenerate configurations of the features (measure-zero events). This is related to the notion of non-accidental features in vision. We illustrate the approach by categorizing motion sequences in a simplified visual domain.