Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Averaged Gradient Episodic Memory (A-GEM)
Improved variant of GEM that solves the quadratic optimization problem by using a more efficient approximation of gradient constraints, reducing computational complexity while preserving performance.
Episodic Memory Replay
Mechanism for selective recall of specific episodes stored in long-term memory, allowing targeted reconsolidation of relevant knowledge when learning new tasks.
Selective Experience Replay
Memorization strategy that adaptively selects experiences to retain based on their relevance, diversity, or expected contribution to avoiding catastrophic forgetting.
Herding Selection
Method for selecting representative examples that maximizes diversity in feature space, choosing points that minimize the average distance to all other points of the same class.
Class-Balanced Sampling
Sampling strategy that ensures fair representation of classes in replay memory, preventing distribution bias and improving generalization on imbalanced tasks.
Meta-Learning with Memory
Learning paradigm where the model learns to learn by using episodic memory to store and retrieve rapid adaptation strategies based on previous learning experiences.
Dynamic Memory Networks
Augmented memory architecture using dynamic attention mechanisms and episodic memory modules to sequentially process information and build rich representations for reasoning.
Episodic Memory Module
Neural component specialized in storing and retrieving specific experiential episodes, implementing content-addressing mechanisms and differentiable updates for continuous learning.