AI 용어집
인공지능 완전 사전
Meta-Learning
Approach where a model learns to learn by training on a variety of different tasks to develop rapid generalization capability to new tasks with few examples. Meta-learning optimizes model initialization parameters to facilitate rapid adaptation.
Vision-Language Models
Multimodal architectures designed to simultaneously understand and generate visual and textual content by learning shared representations between these modalities. These models leverage synergies between vision and language for tasks like image captioning or VQA.
Adaptive Fine-Tuning
Strategy for adapting pre-trained models where only a subset of parameters is updated with few examples, thus preserving general knowledge while specializing rapidly. This approach minimizes the risk of overfitting in few-shot scenarios.
Cross-Modal Similarity Metric
Mathematical function quantifying semantic relevance between elements of different modalities, essential for multimodal alignment and information retrieval. These metrics learn to map heterogeneous spaces to meaningful comparison.
Multimodal Transfer Learning
Application of knowledge acquired from source tasks or modalities to improve performance on a target multimodal task with limited data. This approach leverages inter-task relationships to reduce training data requirements.