AI 용어집
인공지능 완전 사전
Hard Parameter Sharing
Multi-Task Learning approach where the lower layers of the network are shared between all tasks, while only the upper layers are specific to each task.
Soft Parameter Sharing
Technique where each task has its own model with its own parameters, but regularization is applied to encourage similarity between the parameters of models from different tasks.
Cross-Task Regularization
Regularization method that uses knowledge from a source task to constrain and improve learning on a target task, reducing overfitting.
Task-Specific Layers
Neural layers dedicated to a particular task in a multi-task architecture, allowing specialization while benefiting from shared representations of lower layers.
Multi-Head Architecture
Neural network structure with a shared common trunk and multiple specialized prediction heads, each optimized for a different task in a multi-task context.
Shared Representation Learning
Process of learning latent representations that capture features useful simultaneously for multiple tasks, maximizing inter-task knowledge transfer.
Task Relationship Modeling
Technique aimed at quantifying and explicitly exploiting relationships between different learning tasks to optimize representation sharing and improve overall performance.
Progressive Neural Networks
Architecture where new neural columns are added for new tasks while preserving lateral connections to the columns of previous tasks, avoiding catastrophic forgetting.
Transferability Analysis
Quantitative evaluation of the ability of features learned on a source task to be effectively transferred to a different but related target task.
Task Uncertainty Weighting
Multi-task optimization method that automatically weights the loss of each task according to its homoscedastic uncertainty, balancing learning between tasks.
Task Clustering
Approach that groups similar tasks into clusters to optimize the sharing of representations, enabling more effective transfer within groups of related tasks.
Multi-Modal Transfer Learning
Extension of transfer learning where knowledge is transferred between different data modalities (text, image, audio) to enrich shared representations.