Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Local vs Global Attribution
Distinction between methods explaining individual predictions (local) and those evaluating feature importance across the entire dataset (global).
Model-agnostic Attribution
Explanation approaches that work independently of the model's internal architecture, treating it as a black box to generate attributions.
Model-specific Attribution
Attribution techniques that exploit the specific internal structure of a model (decision trees, neural networks) to provide more accurate explanations.
Attention Mechanisms
Neural network components that learn importance weights for different parts of the input, naturally serving as an attribution mechanism.
Layer-wise Relevance Propagation
Technique propagating the prediction backward through the network, distributing output relevance to neurons and input features layer by layer.
Feature Ablation
Systematic technique of removing or masking features to evaluate their individual impact on model performance.
Path Attribution
Attribution method following activation paths in the neural network to assign credit to input features based on their contribution flow.
Gradient-based Attribution
Family of methods using gradients of the output with respect to inputs to quantify feature sensitivity and importance.
Feature Interaction
Measurement of the combined effect of pairs or groups of features on the model's prediction, beyond their individual contributions.
Input Attribution
Process of assigning importance scores to specific input elements (pixels, words, variables) to explain their role in the model's decision.