Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Saliency Map
A heatmap that visualizes the pixels or features of an input that have the greatest impact on a deep learning model's prediction.
Output Gradient
Method that computes the gradient of the model's output with respect to each input pixel to quantify its influence on the final prediction.
Score-CAM
Variant of Grad-CAM that does not use gradients, but evaluates the importance of each activation map by passing its upscaling through the network.
FullGrad
Attribution method that decomposes a neural network's output into contributions from each neuron and bias, providing a complete and gradient-sensitive explanation.
Input-Gradient
Explainability method that uses the gradient of the output with respect to the input as a direct measure of relevance for each pixel or feature.
Perturbation-based Analysis
Family of techniques that evaluate feature importance by introducing systematic modifications to the input data and measuring the effect on the output.
RISE (Randomized Input Sampling for Explanation)
Method that generates explanations by randomly masking the input with multiple masks and aggregating predictions to estimate the importance of each region.