We propose a computational model of the impact of labels on visual categorisation. The proposed model is based on selforganising maps. The model successfully reproduces the experiments demonstrating the impact of labelling on infant categorisation reported in Plunkett, Hu, and Cohen (2008). Two architectures are explored. Both mimic infant behaviour in both the familiarisation and testing phases of the procedure, using a training regime which involves only single presentations of each stimulus. The model reproduced these results in the absence of a explicit teaching signal and predicts that the observed behaviour in infants is due to a transient form of learning that might lead to the emergence of hierarchically organised categorical structure.