AI 용어집
인공지능 완전 사전
Zero-Shot Learning (ZSL)
Machine learning paradigm where the model recognizes classes never seen during training by using auxiliary semantic information such as textual descriptions or attributes.
Semantic Decomposition
Process of decomposing complex classes into fundamental and shared attributes, facilitating generalization to new classes by composing these atomic elements.
Attribute Space
High-dimensional vector space where each dimension represents a specific semantic attribute, serving as a common domain for seen and unseen classes.
Semantic Embedding
Dense vector representation that captures semantic relationships between concepts, enabling similarity measurement and knowledge transfer between classes.
Attribute-Class Mapping
Function learning the correspondence between attribute representations and classes, essential for classification in zero-shot scenarios.
Attribute Inference
Recognition process where the model first detects attributes present in an instance, then infers the probable class based on these observations.
Semantic Knowledge Base
Structure containing semantic relationships between concepts and attributes, exploited as an external knowledge source for zero-shot learning.
Semantic Projection Bias
Phenomenon where ZSL models tend to preferentially predict classes seen during training, even in generalized zero-shot scenarios.
Generalized Zero-Shot Learning (GZSL)
Extension of ZSL where the model must simultaneously classify both seen and unseen classes, better reflecting real-world application scenarios.