Online continual Fréchet distance learning

Task-Free Continual Learning (TFCL) represents a realistic yet challenging learning scenario, aiming to train a model capable of capturing all previously seen important information without forgetting while the task information is not available. Existing studies usually propose learning a fixed-capacity model, leading to a suboptimal performance on complex data. In this research study, we propose a novel scalable mixture model framework to deal with both supervised and unsupervised learning scenarios under TFCL. A Fréchet distance-based model expansion mechanism is proposed to evaluate the knowledge discrepancy between each previously learned component of the mixture and that being currently trained, and use this measure for guiding model expansion. This mechanism is shown to promote the knowledge discrepancy among model's mixture components while enforcing the acquired information diversification, leading to a more compact model. We also propose a novel memory optimization mechanism that utilizes the entire previously learned knowledge to guide how to preserve novel data used for training a new mixture component, further promoting knowledge diversity among model's components. The empirical results, following a series of supervised and unsupervised continual learning experiments, show that the proposed framework achieves state-of-the-art performance.