Please enable JavaScript.
Coggle requires JavaScript to display documents.
Perturbation-based methods - Coggle Diagram
Perturbation-based methods
Statistical approach
Principle
Occlude the image
Use the change in model's output as a measure of importance
Exemples
RISE
Local interpretable model-agnostic explanations (LIME),
Advances/improvement
Other importance measures
Other occlusion strategies
Optimisation approach
Principle
Mask the input image
Find the smallest mask that keeps the output constant
Exemples
Meaningful perturbation (MP), 2017
Extremal perturbations (EP), 2019
Real Time image Saliency
Advances/improvement
regularizing the optimization (mask size, smoothness of the mask)
using integrated gradient for optimization
set certain constraints on the optimization