Black Box Attribution

Black box attribution methods aim to explain the decisions of complex machine learning models, particularly deep neural networks, by identifying which input features most influence the model's output. Current research focuses on improving the interpretability and efficiency of these methods, exploring techniques like those based on dependence measures (e.g., using Hilbert-Schmidt Independence Criterion) and training auxiliary networks to generate more accurate and visually appealing attributions (e.g., generating class-specific masks). These advancements are crucial for building trust in AI systems and enabling better understanding and debugging of complex models across diverse applications, including image classification and object detection.

Papers

November 27, 2022

Latent SHAP: Toward Practical Human-Interpretable Explanations
Ron Bitton, Alon Malach, Amiel Meiseles, Satoru Momiyama, Toshinori Araki, Jun Furukawa, Yuval Elovici, Asaf Shabtai
Fusion SHAP Facial Beauty Prediction Model Agnostic Feature Attribution Black Box Attribution

June 13, 2022

Making Sense of Dependence: Efficient Black-box Explanations Using Dependence Measure
Paul Novello, Thomas Fel, David Vigouroux
Line by Line Explanation Multiple Sens Hilbert Schmidt Independence Criterion Activity Dependence Dependence Measure Black Box Attribution

May 23, 2022

What You See is What You Classify: Black Box Attributions
Steven Stalder, Nathanaël Perraudin, Radhakrishna Achanta, Fernando Perez-Cruz, Michele Volpi
Image Classification Saliency Map Object Mask Agnostic Mask Black Box Attribution

Black Box Attribution

Papers

Latent SHAP: Toward Practical Human-Interpretable Explanations

Making Sense of Dependence: Efficient Black-box Explanations Using Dependence Measure

What You See is What You Classify: Black Box Attributions