keyword
dynamic representation
A dynamic representation is a machine learning feature encoding that adapts, expands, or reorganizes over time as new data, classes, or tasks are introduced, rather than remaining fixed after initial training. In contrast to static embeddings, a dynamic representation allows a model to continuously update its latent feature space or expand its network capacity to incorporate novel patterns while preserving previously acquired knowledge. This capability is essential in continual and incremental learning scenarios, as it balances stability and plasticity to prevent catastrophic forgetting when training on evolving data streams without requiring complete retraining on past data.
1 item
