KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Space embedding
Vector space where visual and semantic features are projected to enable comparisons and associations between seen and unseen classes during training.
Attribute-based learning
Methodology using common descriptive attributes to share information between classes, enabling recognition of unseen objects through their shared characteristics.
Prototype networks
Neural network architecture that learns a metric space where classification is performed by computing distances to class prototypes, which are mean representatives of each category.
Model-Agnostic Meta-Learning
Meta-learning algorithm that optimizes a model's initialization parameters to enable rapid adaptation to new tasks with few gradient updates.
Cross-domain adaptation
Technique that allows a model trained on a source domain to effectively adapt to a different target domain, with potentially dissimilar data distributions.
Open-set recognition
Classification problem where the model must identify not only known classes but also recognize and reject samples belonging to unknown classes.
Transductive learning
Learning paradigm where the model simultaneously uses training and test data to directly optimize its predictions on the specific test set considered.
Episodic training
Training strategy in meta-learning where each episode simulates a few-shot task with support and query sets to learn rapid adaptation capabilities.
Attribute vocabulary
Predefined set of descriptive attributes used to characterize classes and enable the transfer of semantic knowledge between categories in zero-shot systems.