KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Model Uncertainty
Measure quantifying the model's lack of confidence in its predictions, often used as the main criterion for sample selection in active learning.
Sample Pool
Set of unlabeled data from which the active learning algorithm selects the most relevant samples to be annotated by a human oracle.
Kernel Sampling
Active learning strategy that selects samples maximizing diversity in the feature space using kernel methods to avoid redundancy.
Expected Model Error
Advanced sampling criterion estimating the potential reduction in the model's generalization error if a specific sample were added to the training set.
Committee-Based Active Learning
Approach where multiple models (committee) are trained on the same dataset and disagreements between their predictions are used to measure uncertainty and select samples.
Graph-Based Density
Sampling method that considers both uncertainty and sample density in the feature space by constructing a neighborhood graph to avoid outliers.
Training Set Variation
Strategy measuring the impact of adding a sample on the model's parameters, selecting those that cause the greatest change in the weight space.
Semi-Supervised Active Learning
Combination of active learning and semi-supervised learning where the model learns both from selected samples for annotation and from the inherent structure of unlabeled data.
Prediction Entropy
Uncertainty measure based on the entropy of the model's output probability distribution, where predictions with the highest entropy are selected for annotation.
Reinforcement Active Learning
Framework where the sample selection policy is optimized as a reinforcement learning agent, learning to maximize long-term model improvement.
Annotation Cost
Factor accounting for the resources (time, money) required to label a sample, integrated into active learning strategies for realistic optimization.
Query-Based Active Learning
Paradigm where the model explicitly formulates queries to obtain specific information (labels, features) rather than simply selecting complete samples.
Informative Sample Synthesis
Advanced technique where the model generates new synthetic samples in regions of maximum uncertainty, rather than being limited to selection from the existing pool.