AI-ordlista
Den kompletta ordlistan över AI
Continual Learning without Tasks
Learning paradigm where the model learns continuously from a data stream without explicit task delimitation, requiring autonomous adaptation to distribution shifts.
Plasticity-Stability
Fundamental dilemma in continual learning between the ability to learn new information (plasticity) and the preservation of acquired knowledge (stability).
Dynamic Replay Buffer
Adaptive memory buffer that selectively stores representative samples from the past for periodic review of prior knowledge without defined task boundaries.
Continual Synaptic Regularization
Technique preserving important synaptic weights identified dynamically during continual learning to limit catastrophic forgetting without prior knowledge of tasks.
Adaptive Network Expansion
Strategy dynamically adding neurons or layers to the network when new capabilities are required, enabling organic growth without explicit task segmentation.
Online Synaptic Consolidation
Process of progressive stabilization of important neural connections identified in real-time during continual learning, mimicking biological consolidation mechanisms.
Automatic Change Detection
Mechanism automatically identifying transitions in data distribution or concepts without explicit supervision, crucial for continuous adaptation.
Borderless Episodic Memory
Memory system storing significant experiences continuously without task segmentation, using dynamic selection criteria based on utility.
Continuous Self-Supervised Learning
Approach where the model generates its own learning signals from unlabeled data in a continuous manner, adapting its representations without external supervision.
Continual Latent Space
Reduced-dimension representation evolving dynamically to accommodate new concepts while preserving the semantic structure of prior knowledge.
Gradual Knowledge Transfer
Process of automatic identification and reuse of transferable knowledge between emerging concepts in a continuous data stream without explicit delimitation.
Continuity Metrics
Indicators quantifying model performance on the entirety of knowledge acquired over time, measuring the balance between learning and preservation without task segmentation.
Non-Stationary Adaptation
Ability of a system to modify its internal parameters in response to dynamically changing data distributions, an essential characteristic of task-free learning.