when new knowledge needs to be included in a classifier, the model is retrained from scratch using a huge training set that contains all available information of both old and new knowledge. However, in this talk, we present a way to include new information in a previously trained model without training from scratch and using a small subset of old data. We perform a thorough experimental evaluation of the proposed approach on two image classification datasets: CIFAR-100 and ImageNet. The experiment results show that it is possible to include new knowledge in a model without forgetting the previous one, although, the performance is still lower than training from scratch with the complete training set.
Buscar en Dimension
