KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
MAML (Model-Agnostic Meta-Learning)
Meta-learning algorithm that learns initial model parameters enabling fast adaptation to new tasks with few examples through gradient optimization.
Meta-LSTM
LSTM variant that meta-learns its own update parameters, enabling dynamic adaptation of model weights based on task-specific characteristics.
SNAIL (Simple Neural Attentive Learner)
Hybrid architecture combining temporal convolutions and attention mechanisms to learn from example sequences and rapidly adapt to new tasks.
Meta-SGD
Extension of MAML that learns not only initial parameters but also parameter-specific learning rates for more flexible adaptation to new tasks.
Reptile
Simplified meta-learning algorithm that interpolates between initial weights and weights after a few optimization steps on the current task.
TADAM (Task-Dependent Adaptive Metric)
Method combining prototypical networks with a task-aware attention module to dynamically adapt the embedding space based on task characteristics.
LEAP (Learning to Evaluate)
Framework that meta-learns an evaluation function to compare models across different tasks, directly optimizing meta-generalization performance rather than individual task losses.
L2L (Learning to Learn)
Paradigm where a neural meta-optimizer learns to update the parameters of another network, discovering problem-specific adaptive optimization algorithms.
R2D2 (Recursive Reward Decomposition)
Meta-reinforcement learning method using a hierarchical decomposition of rewards to learn reusable policies across different tasks.
Meta-Transfer Learning
Approach combining meta-learning and transfer learning to learn transferable representations while preserving the ability to quickly adapt to new data distributions.
Meta-RL (Meta-Reinforcement Learning)
Field where the agent learns to learn fast adaptation policies for new reinforcement learning tasks by exploiting regularities across environments.