The paper introduces iCaRL (Incremental Classifier and Representation Learning), a novel training strategy designed to enable systems to learn incrementally, adding new classes over time from a stream of data. Unlike previous methods that are limited to fixed data representations, iCaRL learns both strong classifiers and a data representation simultaneously, making it compatible with deep learning architectures. The key components of iCaRL include: 1. **Classification by Nearest-Mean-of-Exemplars**: This rule ensures robustness against changes in the data representation while requiring only a small number of exemplar images per class. 2. **Prioritized Exemplar Selection Based on Herding**: This step selects exemplars dynamically based on their importance, ensuring that the exemplar set remains efficient and effective. 3. **Representation Learning Using Knowledge Distillation and Prototype Rehearsal**: This step updates the feature extraction routine using exemplars and distillation to prevent catastrophic forgetting. Experiments on the CIFAR-100 and ImageNet ILSVRC 2012 datasets demonstrate that iCaRL can learn many classes incrementally over a long period, outperforming other methods that quickly fail. The paper also discusses related work and highlights the importance of exemplar images in preventing catastrophic forgetting. Despite its promising results, the authors note that further improvements are needed to match the performance of batch-trained systems.The paper introduces iCaRL (Incremental Classifier and Representation Learning), a novel training strategy designed to enable systems to learn incrementally, adding new classes over time from a stream of data. Unlike previous methods that are limited to fixed data representations, iCaRL learns both strong classifiers and a data representation simultaneously, making it compatible with deep learning architectures. The key components of iCaRL include: 1. **Classification by Nearest-Mean-of-Exemplars**: This rule ensures robustness against changes in the data representation while requiring only a small number of exemplar images per class. 2. **Prioritized Exemplar Selection Based on Herding**: This step selects exemplars dynamically based on their importance, ensuring that the exemplar set remains efficient and effective. 3. **Representation Learning Using Knowledge Distillation and Prototype Rehearsal**: This step updates the feature extraction routine using exemplars and distillation to prevent catastrophic forgetting. Experiments on the CIFAR-100 and ImageNet ILSVRC 2012 datasets demonstrate that iCaRL can learn many classes incrementally over a long period, outperforming other methods that quickly fail. The paper also discusses related work and highlights the importance of exemplar images in preventing catastrophic forgetting. Despite its promising results, the authors note that further improvements are needed to match the performance of batch-trained systems.