AI 용어집
인공지능 완전 사전
Co-attention Mechanism
Bidirectional attention architecture where two modalities attend to each other simultaneously, enabling symmetrical interaction and cross-understanding of information.
Fusion Attention
Attention technique that dynamically combines representations from different modalities by learning contextual fusion weights for each data point.
Self-Attention Multimodal
Mechanism where each element in a modality calculates its relative importance compared to all other elements, including those from other modalities in the joint space.
Bilinear Attention
Attention method using bilinear transformations to model complex interactions between modality pairs and capture non-linear relationships.
Attention Alignment
Process of semantic alignment between segments of different modalities using attention maps to identify spatial or temporal correspondences.
Modality-specific Attention
Attention mechanism tailored to the intrinsic characteristics of each modality, using distinct parameters to optimize information selection according to data type.
Dynamic Attention Weighting
System for automatically adjusting attention weights in real-time based on contextual relevance and confidence of multimodal information for each input.
Multi-head Cross-modal Attention
Extension of multi-head attention where each head specializes in capturing different types of intermodal relationships for a richer and more diverse representation.
Attention Bottleneck
Attention layer that forces selective compression of multimodal information into a fixed-dimensional vector, preserving only the most relevant features.
Gated Multimodal Attention
Mechanism using learned gates to selectively control information flow between modalities, enabling fine regulation of multimodal integration.
Adaptive Attention Networks
Neural networks that dynamically adjust their attention strategy based on the quality and availability of information from each modality.
Attention Fusion Layer
Specialized layer that combines outputs from multiple multimodal attention mechanisms using learned weights to optimize the final representation.
Sparse Cross-modal Attention
Cross-modal attention variant that focuses only on the most relevant feature subsets, reducing computational complexity while preserving important relationships.
Temporal Multimodal Attention
Attention mechanism specialized in modeling temporal dependencies between synchronized or unsynchronized modalities in sequential data.
Attention-guided Feature Selection
Process where attention weights serve as a guide to dynamically select the most informative features from each modality before fusion.