Staying Consistent with GradCAM
This research project was conducted as part of COMPSCI 682 Neural Networks: A Modern Introduction, under the guidance of Professor Subhransu Maji and Chuang Gan. Thank you to my collaborators Aaron Sun and Phat Nguyen.
Abstract from:
Staying Consistent with GradCAM
Explainable methods, such as GradCAM, are not always consistent across image transformations. They are oftentimes too broad to be a source of reliable explanations. We present an approach to train a Convolution Neural Network to simultaneously produce consistent explanations using saliency maps from post-hoc explanation methods as regularizers. This in turn creates more robust and confident explanations as well as predictions. We show that our method improves the consistency of this explainability method without any significant accuracy tradeoff.