AI 용어집
인공지능 완전 사전
Metric-based Meta-Learning
An approach that learns a distance or similarity metric to compare examples and make predictions on new tasks.
Model-based Meta-Learning
Methods that use models with internal memory or attention mechanisms to quickly adapt to new tasks.
Optimization-based Meta-Learning
Techniques that directly optimize the learning process to enable rapid adaptation with few gradient updates.
MAML (Model-Agnostic Meta-Learning)
Algorithm that trains models with optimal parameter initialization for fast learning on new tasks.
Prototypical Networks
Architecture that learns an embedding space where each class is represented by a prototype computed from support examples.
Siamese Networks
Twin neural networks that learn to measure similarity between pairs of inputs for few-shot learning.
Matching Networks
Models that use weighted attention mechanisms to match test examples to support examples.
Relation Networks
Architecture that learns a relation function to compare support and test example embeddings.
Memory-Augmented Neural Networks
Neural networks with external memory enabling rapid storage and efficient retrieval of information for new tasks.
Meta-Reinforcement Learning
Application of meta-learning to reinforcement learning problems for rapid adaptation to new environments.
Continual Meta-Learning
Approach combining meta-learning and continual learning to continuously learn on new tasks without forgetting previous ones.
Meta-Learning for Hyperparameter Optimization
Using meta-learning to automatically optimize the hyperparameters of learning models.
Neural Architecture Search with Meta-Learning
Application of meta-learning to automatically discover optimal neural network architectures for specific tasks.
Zero-Shot Learning
Ability to recognize classes never seen during training by using semantic information or descriptions.
One-Shot Learning
Subfield of few-shot learning where the model must learn from a single example per class.
Meta-Learning for Few-Shot Classification
Specialization of meta-learning focused on classification problems with very few training examples per class.
Task-Agnostic Meta-Learning
Approach that learns universal representations without prior knowledge of the distribution of future tasks.