Class-Incremental Learning Enhanced: Assessing New Exemplar Selection Strategies in the iCaRL Framework

The field of machine learning has increasingly focused on incremental learning, enabling systems to continually adapt and improve by integrating new knowledge while retaining previously learned information. One particular area of interest is class-incremental learning, where the learning system sequentially acquires new classes without access to or with limited exposure to past data. The primary challenge in class-incremental learning is catastrophic forgetting, wherein the model tends to overlook previously learned classes when confronted with new tasks. One of the class-incremental learning approaches to mitigate catastrophic forgetting is the Incremental Classifier and Representation Learning (iCaRL) framework. In this paper, we propose three new selection criteria for the iCaRL approach. Our best selection criterion, which uses the K-Means clustering algorithm to create diverse groups and then selects exemplars close to the centroids of the clusters, outperforms the original iCaRL selection criterion by over 16% for the MNIST dataset and by over 12% for the FashionMNIST dataset in terms of average accuracy. The full implementation of the iCaRL approach, along with the three proposed selection criteria and detailed experimental results logs, can be found in our publicly available GitHub repository. By contributing to the ongoing development of class-incremental learning techniques, we aim to support the creation of more effective and robust lifelong machine learning systems.