Knowledge Consolidation based Class Incremental Online Learning with Limited Data

We propose a novel approach for class incremental online learning in a\nlimited data setting. This problem setting is challenging because of the\nfollowing constraints: (1) Classes are given incrementally, which necessitates\na class incremental learning approach; (2) Data for each class is given in an\nonline fashion, i.e., each training example is seen only once during training;\n(3) Each class has very few training examples; and (4) We do not use or assume\naccess to any replay/memory to store data from previous classes. Therefore, in\nthis setting, we have to handle twofold problems of catastrophic forgetting and\noverfitting. In our approach, we learn robust representations that are\ngeneralizable across tasks without suffering from the problems of catastrophic\nforgetting and overfitting to accommodate future classes with limited samples.\nOur proposed method leverages the meta-learning framework with knowledge\nconsolidation. The meta-learning framework helps the model for rapid learning\nwhen samples appear in an online fashion. Simultaneously, knowledge\nconsolidation helps to learn a robust representation against forgetting under\nonline updates to facilitate future learning. Our approach significantly\noutperforms other methods on several benchmarks.\n