AI 용어집
인공지능 완전 사전
Prototypical Networks
Few-shot learning architecture that learns a metric space where classes are represented by prototypes calculated as the mean of embeddings of support examples.
Episode Training
Training strategy in few-shot learning where each episode simulates a few-shot task with a support set and a query set to mimic test conditions.
Support Set
Set of labeled examples provided to the model during inference to help it understand and classify new classes with very few available examples.
Query Set
Set of unlabeled examples that the model must classify using knowledge acquired from the support set during few-shot evaluation.
Metric Learning
Machine learning field aiming to learn a distance or similarity function that brings similar examples closer and pushes different ones apart, fundamental in few-shot learning.
One-Shot Learning
Extreme case of few-shot learning where the model must learn to recognize new classes from only one example per class during inference.
Relation Networks
Few-shot architecture that explicitly learns a comparison function to measure the relationship between support and query examples in an embedding space.
Base Classes
Training categories with many available examples used to pre-train the model before few-shot adaptation to new classes.
Novel Classes
New classes with few or no examples that the model must learn to recognize during the test phase in few-shot learning.
Cross-Domain Few-Shot
Variant of few-shot learning where the target classes come from a different domain than the training classes, presenting a more complex transfer challenge.
Feature Embedding
Low-dimensional vector representation of input data that captures essential semantic features, crucial for comparison in few-shot learning.
Matching Networks
Few-shot architecture that uses an attention mechanism to compare each query example with all support examples and generate a weighted prediction.
Task-Agnostic Pretraining
Pre-training phase where the model learns general representations without knowledge of the specific few-shot tasks it will encounter later.
Adaptive Fine-Tuning
Technique for rapid adaptation of model weights with few iterations on support examples to adapt to new classes in few-shot learning.
Hierarchical Few-Shot
Few-shot approach that exploits hierarchical relationships between classes to improve generalization when few examples are available.
Self-Supervised Few-Shot
Combination of self-supervised learning with few-shot learning to improve representations before adapting to new classes with few examples.