Home // PATTERNS 2022, The Fourteenth International Conference on Pervasive Patterns and Applications // View article

Contributions by Feature Layers in Two-Class Deep LearningImage Classification Decisions

Authors:
Debanjali Banerjee
Chee-Hung Henry Chu

Keywords: Deep Learning, Explainable Artificial Intelligence, Shapley Values, Image Classification

Abstract:
Deep learning methods have excellent accuracy achievements in image classification but largely remains a black box method. Image classification is the core of many machine vision tasks, including object detection. Better understanding of how the classification decision is made will improve the understanding of such tasks as object detection. In this work, we train a deep learning network to classify between two classes. We compute the so-called SHapley Additive exPlanations (SHAP) values for the feature layers using input images against a population of other training images for the classification layer. The SHAP value is a special case of the Shapley value which explains the factors in a machine learning decision by measuring the output change due to change in each factor. The SHAP value is the Shapley value satisfying local accuracy, missingness, and consistency properties. Experimental results show the different responses from the lowest to the highest feature extraction layers.

Pages: 31 to 36

Copyright: Copyright (c) IARIA, 2022

Publication date: April 24, 2022

Published in: conference

ISSN: 2308-3557

ISBN: 978-1-61208-953-9

Location: Barcelona, Spain

Dates: from April 24, 2022 to April 28, 2022