Semi-supervised learning in infancy: Infants integrate labeled and unlabeled exemplars to learn new categories

AbstractLabels facilitate infants’ category learning. Providing the same label for a set of distinct individuals enhances infants’ ability to identify the underlying category. In infants’ daily life, however, many category exemplars will go unlabeled, and infants will inevitably receive a mix of labeled and unlabeled exemplars when learning real-world object categories. Here, we ask whether 2-year-old infants can integrate these labeled and unlabeled exemplars when learning a novel category. To do so, we draw on machine learning research in semi-supervised learning, a class of algorithms designed to learn from just such mixed data. Our results suggest infants do engage in semi-supervised category learning. Infants learned categories as successfully in a semi-supervised condition as in a fully-labeled condition—and more successfully than in an unlabeled condition. These findings reveal that the power of labels extends beyond the exemplars being labeled: labeling also promotes infants’ learning from subsequent, unlabeled exemplars.

Return to previous page