A Novel Explainable Deep Learning Model with Class Specific Features

Abstract

AbstractThe predictive accuracy of any machine learning model is highly depended on the features used to train the model. For this reason, it is important to extract good discriminative features from the raw data. This extraction of good features from raw data is a challenging task. Deep learning models like Convolutional Neural Networks (CNNs) have the ability to automatically extract features from raw data and also have excellent predictive capabilities. This excellent predictive capability and ability to extract good features from raw data have made CNN very popular, especially in the field of computer vision. Even though CNN is popular, like many other deep learning models it is also notoriously black-box model. The predictions made by a CNN model cannot be explained based on features that influenced the given predictions. In our work we put forth an architecture that has convolutional layers to extract features automatically and the predictions made by this model can be explained based on specific features/neurons that resulted in the prediction. The model put forth in this paper has accuracy that is on par with the state-of-the-art models. Also, its predictions are explainable with target class specific feature importance.KeywordsFeature extractionExplainable deep learningConvolutional neural networksLogistic regression