Eigen-CAM: Visual Explanations for Deep Convolutional Neural Networks

Abstract

The adoption of deep convolutional neural networks (CNN) is growing exponentially in wide varieties of applications due to exceptional performance that equals to or is better than classical machine learning as well as a human. However, such models are difficult to interpret, susceptible to overfit, and hard to decode failure. An increasing body of literature, such as class activation map (CAM), focused on understanding what representations or features a model learned from the data. This paper presents novel Eigen-CAM to enhance explanations of CNN predictions by visualizing principal components of learned representations from convolutional layers. The Eigen-CAM is intuitive, easy to use, computationally efficient, and does not require correct classification by the model. Eigen-CAM can work with all CNN models without the need to modify layers or retrain models. For the task of generating a visual explanation of CNN predictions, compared to state-of-the-art methods, Eigen-CAM is more consistent, class discriminative, and robust against classification errors made by dense layers. Empirical analyses and comparison with the best state-of-the-art methods show up to 12% improvement in weakly-supervised object localization, an average of 13% improvement in weakly-supervised segmentation, and at least 15% improvement in generic object proposal.

Publication Title

SN Computer Science

Link to Full Text

Share

COinS