AI Glossary
The complete dictionary of Artificial Intelligence
Active Learning in Stream
Learning paradigm where the model intelligently selects the most informative data instances in a continuous stream to request their labeling, thus optimizing the performance/cost ratio.
Uncertainty Selection
Active learning strategy that prioritizes instances for which the model exhibits the greatest prediction uncertainty, typically measured by entropy or confidence margin.
Labeling Budget
Quantitative constraint defining the maximum number of labels that can be requested per unit of time or for a given volume of data in a streaming context.
Concept Drift
Non-stationary change in the underlying data distribution or the relationship between features and targets, requiring continuous model adaptation in stream.
Diversity Selection
Instance selection approach aiming to maximize the diversity of labeled examples by avoiding informational redundancy in the feature space.
Human Oracle
External expertise source (typically human) solicited to provide labels for instances selected by the active learning system in real-time.
Labeling Cost
Temporal, financial, or computational resources required to obtain a ground truth label, optimized by active learning strategies.
Representativeness Selection
Instance selection method based on their ability to represent the global or local structure of the data stream, ensuring balanced coverage of the space.
Hybrid Selection Strategy
Optimized combination of multiple selection criteria (uncertainty, diversity, density) to improve the efficiency of active learning in streams.
Annotation Latency
Time delay between the selection of an instance by the model and the receipt of its label, directly impacting real-time performance.
Density-Based Selection
Selection criterion favoring instances located in high data density regions to maximize the informational impact of each label.
Dynamic Adaptation
System's ability to automatically adjust its selection strategy based on detected changes in the stream and resource constraints.
Reservoir Sampling
Algorithm that maintains a fixed-size random sample from a potentially infinite data stream with uniform probability.
Confidence Margin
Uncertainty measure calculated as the difference between the probabilities of the two most probable classes, used to guide active selection.
Novelty Detection
Process of identifying instances or patterns significantly different from previously observed data in the stream, requiring special attention.
Bandwidth Constraints
Limitations on the volume of data that can be transmitted or processed simultaneously, influencing selection decisions in active learning.
Adaptive Stopping Strategy
Mechanism dynamically determining when to stop soliciting labels based on the evolution of model performance and the remaining budget.
Active Collaborative Filtering
Application of active learning to streaming recommendation systems, where user interactions are selectively sampled for learning.
Balancing Exploration-Exploitation
Fundamental dilemma involving the trade-off between exploring new regions of the space and exploiting acquired knowledge in active selection.