iCaRL: Incremental Classifier and Representation Learning

iCaRL targets a core problem on the road to artificial intelligence: incrementally learning systems that acquire more and more concepts over time from a stream of data. The training strategy operates in a class-incremental way in which only the training data for a small number of classes must be present simultaneously and new classes can be added progressively. Crucially, iCaRL learns strong classifiers and a data representation at the same time, distinguishing it from earlier approaches that were limited to fixed data representations and therefore incompatible with deep learning architectures.

Experiments on CIFAR-100 and ImageNet ILSVRC 2012 demonstrate that iCaRL can learn many classes incrementally over a long period of time, whereas other strategies quickly fail under the same conditions. By learning classifiers and a data representation simultaneously, iCaRL overcomes the fixed-representation limitation of earlier methods and stays compatible with deep architectures, enabling systems that accumulate more and more concepts over time as new data arrives.