Output-targeted baseline for neuron attribution calculation

Attribution methods explaining a particular decision for a given convolutional neural network (CNN) have gained a lot of attention over the last few years. Among them, approximation methods of Shapley values are considered to be better ways of assigning attribution scores such that several desirable axioms are satisfied. Nevertheless, these attribution scores may still be misleading or inaccurate due to the inappropriate selection of a baseline which is necessary to apply Shapley values to CNNs. Previous baseline studies have focused on developing a ge-neric baseline selection method for all approximation methods; however, we find that designing a baseline under the essence of the selected approximation method itself produces better results than generic ones. With this ob-servation, we propose two primal baseline properties for Aumann-Shapley-based attributions and design a gen-eral objective function of generating a baseline iteratively by gradient descent. To increase efficiency, we further reduce the objective function into a quadratic optimization problem where the gradients only need to be calcu-lated once. We show that our method produces better attribution results than several state-of-the-art baseline selections and attribution methods on both qualitative and quantitative experiments.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper presents attribution methods for explaining decisions made by convolutional neural networks (CNNs), and proposes an objective function for generating baselines iteratively using gradient descent. The research shows that this method outperforms other approaches in terms of attribution results.