Continual Learning Approaches for Imbalanced Class Distribution Tasks

• 1、North China University of Technology,Beijing 100144

Abstract：Continual learning(CL) enables models to gradually adapt to new tasks and data while retaining previous knowledge, making it particularly suitable for scenarios with dynamic data distributions and limited resources. However, in practical applications, the issue of imbalanced data distributions exacerbates catastrophic forgetting in continual learning, impairing the model\'s ability to learn minority classes and its overall generalization performance. To address this challenge, this paper proposes a continual learning method specifically designed for tasks with imbalanced class distributions,called Continual Learning Methods for Tasks with Imbalanced Class Distributions(ConIm). ConIm combines adaptive adapter routing with an online prototype learning model. The adaptive adapter routing mechanism dynamically assigns dedicated learning paths for different classes, enhancing the model\'s adaptability in representing minority classes. The online prototype learning model prioritizes updating parameters relevant to the current task while protecting critical parameters of previous tasks through an importance-weighted strategy,achieving efficient knowledge integration. Experimental results demonstrate that ConIm outperforms existing continual learning methods on imbalanced datasets, effectively balancing the acquisition of new knowledge and the retention of old knowledge, and significantly improves the model\'s generalization performance in multi-task scenarios.

Keywords： Machine Learning Imbalanced Class Datasets Adaptive Adapter Routing Online Prototype Learning Model