AI-ordlista
Den kompletta ordlistan över AI
Prototype Network
Meta-learning architecture that creates class prototypes by computing the mean of training example embeddings to classify new instances by minimum Euclidean distance.
Class Prototype
Central vector representation of a class calculated as the mean of support set example embeddings, serving as a reference for classification.
Softmax on Distances
Activation function that converts negative distances between queries and prototypes into classification probabilities using the formula exp(-d) / Σ exp(-d).
Prototypical Loss
Loss function optimizing the encoder to minimize intra-class distances and maximize inter-class distances in the prototype space.
Prototype Vectorization
Process of transforming support examples into unique numerical vectors representing the essential features of each class.
Class Separation
Main objective of prototype networks aiming to maximize the distance between prototypes of different classes in the embedding space.
Distance-Based Classification
Alternative approach to parameterized classifiers where predictions are based on metric proximity to class representations rather than learned weights.
Class Centroid
Mathematical point representing the center of gravity of a class's embeddings, calculated as the vector mean to form the prototype.
Embedding Dimension
Size of the vector space in which prototypes and queries are represented, affecting the model's discrimination capability.
Similarity Function
Mathematical function quantifying the proximity between two vectors, used to compare queries to prototypes in prototype networks.
Episode-based Learning
Training strategy where each batch constitutes a complete few-shot task, enabling the model to learn rapid generalization capabilities.
Prototypical Representation
Classification model where each class is represented by a single vector capturing the essential characteristics of the available examples.