KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Synaptic Consolidation
Neurobiologically-inspired process that stabilizes important synaptic connections in artificial neural networks, simulating memory consolidation in the human brain.
Weight Importance Scoring
Quantitative evaluation method of the importance of each neural weight based on their contribution to performance on previous tasks, used to guide regularization.
Task-specific Parameters
Subset of model parameters specialized in performing specific tasks, identified and protected by EWC when learning new skills.
Online EWC
Adaptive variant of EWC that continuously updates the Fisher information matrix and regularization constraints, enabling better adaptation to real-time data streams.
Elastic Regularization
Quadratic penalty term in the EWC loss function that elastically constrains modifications of important weights, thereby preserving learned knowledge.
Memory Replay Buffer
Mechanism complementary to EWC that stores and reuses samples from previous tasks to reinforce knowledge retention against catastrophic forgetting.