keyword
dynamic structure reorganization
Dynamic structure reorganization is a machine learning process in which an artificial neural network dynamically alters, expands, or reconfigures its internal architecture during training to learn new information while preserving previously acquired knowledge. In continual and incremental learning environments, standard models with fixed architectures often suffer from catastrophic forgetting when updated solely on new data streams. Dynamic structure reorganization mitigates this issue by adaptively adjusting the network topology, such as by expanding feature branches, integrating auxiliary pathways, and selectively updating or freezing specific sub-networks. This adaptive mechanism enables the model to balance plasticity for learning novel classes or tasks with stability for retaining historical feature representations without requiring retraining on old data.
1 item
