Słownik AI
Kompletny słownik sztucznej inteligencji
Incremental Learning
Learning paradigm where a model continuously improves from new data without requiring complete retraining on historical data. This approach allows real-time adaptation while preserving previously acquired knowledge.
Continual Learning
AI field aiming to develop systems capable of sequentially learning multiple tasks without complete reset. The goal is to create adaptive models that accumulate knowledge over the long term.
Dynamic Expandable Networks
Models capable of dynamically adapting their architecture by adding new units or layers when new classes or tasks appear. This controlled expansion enables efficient growth without compromising existing performance.
Regularization-based Methods
Family of approaches using penalty terms in the loss function to constrain modifications of important weights. These methods identify and protect critical parameters for performance on previous tasks.
Architecture-based Methods
Continual learning strategies modifying the network structure to accommodate new knowledge without interfering with old knowledge. These approaches include dynamic expansion and dedicated resource allocation per task.
Replay-based Methods
Continual learning techniques using storage and selective reuse of past data to maintain performance. These methods vary in their strategy for selecting and resampling stored experiences.
Task-incremental Learning
Continual learning scenario where task identity is known during inference, allowing the use of masks or specific sub-networks. This simplification facilitates knowledge separation between different tasks.
Domain-incremental Learning
Paradigm where classes remain constant but data distribution changes progressively between tasks. The model must adapt to new domains while maintaining its ability to recognize all classes.
Class-incremental learning
Most challenging scenario where new classes are introduced progressively without knowledge of task identity. The model must distinguish between old and new classes while avoiding forgetting past knowledge.