AI-ordlista
Den kompletta ordlistan över AI
Semantic Embedding
Dense vector representation of classes in a semantic space where relationships between concepts are preserved, allowing the model to understand similarities between unseen classes.
Attribute-Based Classification
Approach where objects are described by a set of interpretable semantic attributes, enabling classification of new classes by combining these predefined attributes.
Seen Classes
Set of object categories used during the model's training phase, for which visual examples are available for learning representations.
Unseen Classes
Object categories absent from training data but that the model must identify during inference using only semantic descriptions or attributes.
Visual-Semantic Mapping
Learning function that projects visual features of images into the semantic space where textual descriptions of classes reside.
Compatibility Function
Function measuring the similarity between a visual representation and a semantic representation, used to assign the most probable class to a given instance.
Generalized Zero-Shot Learning
Extension of ZSL where the model must classify both seen and unseen classes, introducing a classification bias toward familiar classes encountered during training.
Transductive Zero-Shot Learning
ZSL variant where the model has access to test data from unseen classes during training, but without their labels, allowing better alignment of spaces.
Inductive Zero-Shot Learning
Classic ZSL approach where the model learns only from seen classes and must completely generalize to unseen classes without access to test data.
Cross-Modal Transfer
Transfer of knowledge between different data modalities, such as from text to images, essential for ZSL where textual descriptions guide visual recognition.
Semantic Space
High-dimensional vector space where concepts and classes are represented, preserving semantic relationships and enabling reasoning about conceptual similarities.
Label Embedding
Technique transforming class labels into continuous vectors in a semantic space, allowing models to treat classification outputs as regression problems.
Feature Projection
Mathematical operation transforming visual features extracted from images to align them with semantic representations of classes in a common space.
Zero-Shot Domain Adaptation
Combination of ZSL and domain adaptation where the model must classify unseen classes in a target domain different from the training data domain.
Hierarchical Zero-Shot Learning
Approach leveraging the hierarchical structure of classes to improve generalization to unseen categories using parent-child relationships between concepts.
Generative Zero-Shot Learning
Method using generative models to synthesize visual features of unseen classes, transforming ZSL into a classical supervised classification problem.