Maximizing Invariant Data Perturbation with Stochastic Optimization
This work addresses the challenge of generating accurate saliency maps for explaining deep neural network decisions, which is important for improving interpretability in AI applications, but it is incremental as it builds on existing perturbation-based methods.
The paper tackles the difficulty of solving the optimization problem in perturbation-based feature attribution methods by reformulating it as a differentiable function maximization, enabling the use of stochastic gradient-based algorithms like SGD, RMSProp, and Adam, and shows effective identification of relevant image parts in experiments with VGG16 on image classification.
Feature attribution methods, or saliency maps, are one of the most popular approaches for explaining the decisions of complex machine learning models such as deep neural networks. In this study, we propose a stochastic optimization approach for the perturbation-based feature attribution method. While the original optimization problem of the perturbation-based feature attribution is difficult to solve because of the complex constraints, we propose to reformulate the problem as the maximization of a differentiable function, which can be solved using gradient-based algorithms. In particular, stochastic optimization is well-suited for the proposed reformulation, and we can solve the problem using popular algorithms such as SGD, RMSProp, and Adam. The experiment on the image classification with VGG16 shows that the proposed method could identify relevant parts of the images effectively.