Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Domain Shift
Change in data distribution between the training domain (source) and test domain (target), requiring specific adaptation techniques.
Covariate Shift
Specific type of domain shift where the distribution of input features changes between source and target domains, but the conditional distribution P(y|x) remains unchanged.
Domain Discrepancy
Quantitative measure of the difference between data distributions of source and target domains, often used to assess the difficulty of adaptation.
Feature Alignment
Process of aligning feature representations between source and target domains to reduce distributional differences and improve knowledge transfer.
Adversarial Domain Adaptation
Approach using adversarial networks to learn domain-invariant representations by minimizing the ability of a discriminator to distinguish between source and target domains.
Maximum Mean Discrepancy
Statistical metric measuring the distance between distributions of two domains by comparing means in an RKHS kernel feature space.
Domain Invariant Features
Feature representations that remain stable and discriminative across different domains, enabling effective model generalization.
Unsupervised Domain Adaptation
Domain adaptation where no labels are available in the target domain, requiring self-supervised or data structure-based methods.
Supervised Domain Adaptation
Domain adaptation using a limited set of labels in the target domain to guide the alignment process and improve model performance.
Domain Generalization
Extension of domain adaptation aiming to create models that perform well on unseen domains during training, without access to target data.
Multi-source Domain Adaptation
Adaptation using multiple source domains to improve robustness and generalization capability to a single target domain.
Domain Confusion Loss
Loss function designed to maximize the uncertainty of a domain classifier, thereby encouraging the learning of domain-invariant features.
Distribution Matching
Technique aiming to minimize the divergence between source and target data distributions through statistical alignment or feature transformation.
Target Domain Sampling
Strategy for selecting representative samples from the target domain to optimize adaptation efficiency and reduce computational requirements.
Domain-aware Training
Training paradigm that explicitly incorporates domain information into the learning process to improve model adaptability.
Progressive Domain Adaptation
Sequential adaptation method where the model progressively adjusts through intermediate domain steps to facilitate the transition.
Domain-specific Batch Normalization
Normalization technique using distinct statistics for each domain, allowing the model to adapt to distributional variations.