AI 용어집
인공지능 완전 사전
iCaRL (incremental Classifier and Representation Learning)
Incremental learning framework combining example replay with knowledge distillation techniques to maintain performance on previously learned classes.
Dark Experience Replay (DER)
Replay approach that stores not only raw data but also latent representations and logits for more effective reconstruction of past knowledge during continual learning.
Hindsight Replay
Selective replay strategy that uses meta-information about past performance to identify and reuse the most critical examples to prevent catastrophic forgetting.
Reservoir Sampling
Random sampling algorithm that maintains a fixed-size representative set from a data stream, guaranteeing each example an equal probability of being kept in the replay buffer.
Ring Buffer Replay
Circular memory structure that replaces old examples with new ones when the buffer is full, favoring the most recent data while maintaining constant memory size.
Balanced Replay
Replay technique that maintains a balance between examples from different classes or tasks to avoid distribution bias and ensure equitable knowledge distribution.
Curriculum Replay
Organized replay approach that presents old examples according to an optimal pedagogical sequence, typically from simple to complex, to facilitate integration of new knowledge.
Hybrid Replay
Strategy combining multiple replay methods (raw data, generated samples, representations) to maximize knowledge retention while optimizing memory resource usage.
Temporal Replay
Méthode de replay qui considère la séquence temporelle des exemples, privilégiant la rétention des patterns dépendants du temps et des relations causales dans les données séquentielles.
Coreset Selection Replay
Méthode algorithmique qui sélectionne un sous-ensemble minimal d'exemples (coreset) maximisant la représentativité des données passées pour un budget mémoire donné.
Neural Episodic Control
Architecture combinant mémoire épisodique différentiable et replay sélectif pour un apprentissage continu efficace, particulièrement adaptée aux environnements de renforcement.
Meta-Learning Replay
Approche où le modèle apprend méta-comment sélectionner et utiliser efficacement les exemples de replay, s'adaptant dynamiquement aux caractéristiques des nouvelles tâches.