FRMM: Feature Reprojection for Exemplar-Free Class-Incremental Learning

Class-incremental learning involves handling a sequence of classification tasks arriving chronologically, with each task containing different classes. The model must gradually expand its knowledge by learning these tasks successively to acquire the ability to classify all encountered classes. Unlike exemplar-based class-incremental learning (EBCIL) which allows storing some old samples, exemplar-free class-incremental learning (EFCIL) faces a more severe forgetting problem due to the complete prohibition on accessing old data. Some previous methods freeze the feature extractor after the initial stage to enhance the model's stability on the base classes, but this also leads to a larger distribution discrepancy between the feature vectors of base and incremental classes. We propose to insert a reprojection layer between the feature extractor and the classifier to project feature vectors onto a unified distribution, facilitating subsequent classification. Feature reprojection improves the performance of both linear and nearest class mean (NCM) classifier, but their outputs exhibit different biases. We treat them as different experts by ensembling their predictions to combine their strengths. Additionally, we introduce a momentum update to alleviate the linear classifier's inherent forgetting. Across three public datasets of varying scales, our method, named FRMM, outperforms others in most cases while requiring only 1/256 of the storage by the state-of-the-art method that saves a covariance matrix for each class. Code is available at https://github.com/BlacknoSheep/CIL.