Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
N-pair Loss
Generalization of triplet loss that compares one positive example to N negative examples simultaneously, improving learning efficiency and convergence stability. It is particularly effective for large-scale datasets.
Angular Loss
Loss function that optimizes the angles between embedding vectors rather than their Euclidean distances, offering better scale invariance and improved class separation. It is particularly useful for facial recognition tasks.
Center Loss
Loss function that minimizes intra-class distance by forcing embeddings of the same class to approach their learned class center. It is often combined with other discriminative losses to improve cluster compactness.
ArcFace Loss
Additive angular margin loss that improves embedding discriminability by adding an angular margin on hypersphere space. It is particularly effective for recognition tasks with many classes and few examples.
SphereFace Loss
Multiplicative angular loss function that imposes angular constraints on embeddings to improve inter-class separability. It is designed to learn discriminative embeddings on a unit hypersphere.
Proxy-NCA Loss
Variant of Neighborhood Components Analysis that uses learnable proxies to represent each class, reducing computational complexity compared to pair-based approaches. It is particularly effective for large-scale datasets.
Deep Metric Learning
Application of deep neural networks to learn complex distance or similarity functions from raw data. This approach combines the representational power of deep networks with metric learning principles.
Similarity Learning
Learning paradigm where the model learns to evaluate similarity between pairs of examples rather than directly predicting labels. This approach is fundamental for few-shot learning and recommendation systems.