Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Shared Representation Space
Common vector space where different modalities (text, image, sound) are projected to enable their direct comparison and joint manipulation within the same semantic framework.
Multimodal Alignment
Learning process that establishes semantic correspondences between elements of different modalities by minimizing their distance in the shared representation space.
Semantic Projection
Mathematical operation transforming modality-specific representations into vectors in a common semantic space while preserving their intrinsic meaning.
Multimodal Embedding
Dense vector representation that simultaneously encodes features from multiple modalities in a continuous, reduced-dimensional space.
Semantic Bridge
Neural architecture establishing explicit connections between modal encoders to facilitate semantic information transfer and representation alignment.
Common Latent Space
Hidden dimension where latent variables from different modalities are constrained to share a common statistical structure to enable their mutual generation.
Modal Synchronization
Process of temporal and semantic alignment of multimodal streams to ensure coherence of shared representations at varying temporal resolutions.
Modal Mapping
Non-linear transformation learning to project modality-specific feature spaces into a common geometry that preserves semantic relationships.
Intermodal Translation
Ability of a multimodal system to generate one modality from another through their shared representation in the common space.
Modal Coherence
Measure quantifying the degree of semantic alignment between representations of different modalities in the shared space, crucial for effective fusions.
Modal Coupling
Regularizing constraint forcing representations of different modalities to maintain consistent structural relationships in the shared space.
Joint Encoding
Strategy where a single encoder simultaneously processes all modalities to directly produce a unified representation rather than separate encodings.
Unified Representation
Single vector synthetically capturing semantic information from all input modalities, enabling homogeneous classification or generation tasks.
Shared Embedding Space
Fixed-dimensional vector domain where embeddings from different modalities coexist and can be compared using standard similarity metrics.
Deep Semantic Alignment
Alignment technique operating at the level of abstract concepts rather than superficial features, using deep architectures to capture relationships.
Feature-Level Fusion
Multimodal fusion approach combining intermediate representations from encoders before their final projection into the shared space.