Answers are grounded in this paper and cite their sources.
This paper studies interpretability of convolutional networks by means of maps. Most approaches based on Class Activation Maps (CAM) combine from fully connected layers and gradient through variants of. However, it is well understood that gradients are noisy and like guided backpropagation have been proposed to obtain better at inference. In this work, we present a novel training approach improve the quality of gradients for interpretability. In particular, we a regularization loss such that the gradient with respect to the image obtained by standard backpropagation is similar to the gradient by guided backpropagation. We find that the resulting gradient is less noisy and improves quantitatively the interpretability of different networks, using several interpretability methods.
No takes yet. Share an insight, caveat, or question.
Figueroa et al. (2024) studied this question.