keyword
self-sustaining representation expansion
Self-sustaining representation expansion is a continual machine learning framework designed to adaptively incorporate new classes into a model without relying on stored historical training exemplars. In this approach, a neural network expands its underlying feature space and adjusts its architectural components to learn novel concepts while concurrently safeguarding previously acquired knowledge. By utilizing structural expansion alongside self-contained knowledge transfer mechanisms, such as selective distillation and prototype-based feature retention, the framework maintains the discriminability of old representations and mitigates catastrophic forgetting using only current training data.
1 item
