KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Model-Agnostic Meta-Learning (MAML)
Meta-learning algorithm that optimizes the initial parameters of a model to enable rapid adaptation with few gradient steps on new tasks.
Task Distribution
Probabilistic distribution defining the set of tasks on which the agent meta-learns, essential for generalization to new similar tasks.
Inner Loop Optimization
Process of rapid adaptation of model parameters on a specific task using a few gradient steps during meta-learning.
Outer Loop Optimization
Update of meta-parameters by aggregating gradients from multiple tasks to improve the overall adaptation capability of the model.
Task Embedding
Compact vector representation of a task learned by the meta-learner to facilitate quick recognition and adaptation to similar tasks.
Meta-Gradient
Gradient calculated through the inner optimization process to update the model's meta-parameters in meta-learning algorithms.
Fast Adaptation
Goal of meta-learning that allows an agent to achieve optimal performance on a new task with minimum interactions.
Meta-Training
Training phase where the agent learns on a distribution of tasks to develop general adaptation capabilities.
Meta-Testing
Evaluation phase where the agent is tested on novel tasks to measure its generalization and rapid adaptation capabilities.
Episode-based Meta-RL
Meta-RL approach where learning is structured in episodes containing training and testing phases for each task.
Hierarchical Meta-Learning
Extension of meta-learning where meta-parameters are organized in multiple hierarchical levels for progressive adaptation to tasks.
Meta-Policy
Learned policy that generates or adapts other task-specific policies, rather than directly controlling the agent.
Contextual Meta-Learning
Variant of meta-learning where the context or task identity is explicitly provided to guide rapid adaptation.
Online Meta-Learning
Approach where meta-learning occurs continuously during interaction with the environment, without clear separation of meta-training phases.
Meta-Exploration
Exploration strategy optimized to quickly discover the characteristics of a new task by leveraging meta-learned knowledge.
Task Uncertainty
Uncertainty about the exact nature of a new task that meta-learning agents must handle to effectively adapt their behavior.
Meta-Regularization
Regularization techniques applied during meta-learning to improve robustness and generalization to new tasks.
Multi-Task Meta-Learning
Extension of meta-learning where the agent must simultaneously adapt to multiple related tasks by effectively sharing knowledge.