Comparing backpropagation and perturbation XAI
After reviewing backpropagation and perturbation, let’s compare the pros and cons of these XAI techniques.
Generally, backpropagation is more efficient than perturbation-based XAI methods in generating importance scores for all input features in a single backward pass through the network. However, backpropagation-based XAI methods are typically prone to noise and require internal information about the model. While feature attributions associated with an area of input image seem intuitive, neighboring individual pixels can experience a high variant of attribution assignments, which is a shortcoming of backpropagation-based methods. Furthermore, a lack of granular description of a model’s characteristics might result in inconsistent correlation to its output variation, resulting in less faithful explanations.
In contrast, perturbation-based methods are widely applicable to any deep learning model, regardless of its architecture...