Signature Activation: A Sparse Signal View for Holistic Saliency

Jose Roberto Tello Ayala, Akl C. Fahed, Weiwei Pan, Eugene V. Pomerantsev, Patrick Thomas Ellinor, Anthony Philippakis, Finale Doshi-Velez

Keywords: Interpretability, model explanation, saliency, healthcare

TL;DR: A new saliency method is introduce that takes into account all the areas of an image that are of relevance for the output of Convolutional Neural Network models.

Abstract: The adoption of machine learning in healthcare calls for model transparency and explainability. In this work, we introduce Signature Activation, a saliency method that generates holistic and class-agnostic explanations for Convolutional Neural Networks' outputs. We exploit the sparsity of images and give theoretical explanation to justify our methods. We show the potential use of our method in clinical settings through evaluating its efficacy for aiding the detection of lesions in Coronary Angiorams.

Submission Number: 93