AI 용어집
인공지능 완전 사전
Multimodal attention
Mechanism allowing the model to dynamically weight the importance of different modalities according to context and task.
Cross-encoding
Process where information from one modality influences the encoding of another modality to create interdependent representations.
Latent space projection
Transformation of multimodal features into a common vector space where they can be compared and combined.
Modal alignment
Technique aimed at semantically matching elements from different modalities to facilitate their integration.
Co-training
Learning method where multiple classifiers trained on different modalities teach each other.
Multimodal transformer
Architecture based on attention mechanisms designed to simultaneously process multiple types of heterogeneous data.
Decision fusion
Integration at the output level of multiple specialized models to produce a consolidated final prediction.
Adaptive modality weighting
Mechanism dynamically adjusting the influence of each modality according to their relevance for the considered input.
Graph-based fusion
Approach using graph structures to model relationships between different modalities and their interactions.
Multimodal contrastive learning
Self-supervised technique learning representations by bringing positive multimodal pairs closer and pushing negative ones apart.
Hierarchical fusion
Strategy organizing the integration of modalities across multiple levels, from local features to global representations.
Modal gating
Mechanism selectively controlling the flow of information between modalities based on their contextual relevance.
Tensor fusion
Mathematical method combining modalities through tensor operations to capture higher-order interactions.
Multimodal co-attention
Bidirectional mechanism where each modality generates attention weights based on the representations of other modalities.