How Can We Promote Generalization of Novels Concepts? Jean-Pierre Thibaut, Luc Augier jean-pierre.thibaut@u-bourgogne.fr; luc.augier@u-bourgogne.fr Université de Bourgogne (France) Abstract One major issue in education is generalization. In many learning situations, participants do not generalize beyond the situations in which they were trained or generalize to situations that are quite similar to the training phase. The purpose of the present contribution is to study conditions that would promote generalization in new concept learning situations. When subjects are confronted with new stimuli and learn to categorize them according to a rule, they have to segment the stimuli into relevant features for categorization. Our central question is to address the question of “history of categorization” that is whether a simplified version will help children to generalize to perceptually more "complex" versions of the relevant features they would be unable to learn with no pre-training with the simpler version. The results indicate that variability must be included in any model of concept learning. First, the probability that a relevant dimension will be discovered depends on the presence and the structure of the other dimensions (irrelevant) that compose the stimuli and, more specifically, that participate in the manifestation of the rule. Second, in order to understand whether or not a particular instance of a dimension will be discovered by children, one has to include the history of categorization of the participants [5]. More generally, we want to argue for a conception of learning which promotes comparisons between instances and which conceives generalization as a constructive process rather than as an automatic process. Introduction During their development, children have to learn to categorize stimuli according to adults' standards. In the course of learning, students are confronted to the same purpose, at any age. In general, learning situations involve the necessity to learn a specified concept. This implies at least two things. The first is that children have to learn the concept. This can be established in different ways. The second is that they should be able to generalize it in new situations. In the case of categorization, learning a new concept should mean that they have to find the relevant features for categorization and apply it to new objects. If their particular task is to learn to categorize a set of new objects into two new categories, they will have to find the features that characterize the stimuli of each category and that distinguish them from stimuli of the other category. Thus, children, like adults, are presumed to formulate and test simple hypotheses concerning the rule that define membership. This means that participants will analyze stimuli into their dimensions and test whether each dimension partitions the set of stimuli. Various dimensions of the stimuli contribute to the difficulty of the task such as salience of dimensions: a non salient relevant dimension among salient irrelevant dimensions presumably requires more systematic analyses of the stimuli than a salient relevant dimension among non salient irrelevant dimensions. Variability in the perceptual manifestation of a relevant feature can hinder this relevant feature and impede its discovery. For example, compare Figure 1A stimuli with Figure 1B stimuli which define two experimental conditions. In the two conditions, the stimuli come from two categories defined by the same