Argument based machine learning (ABML) is an extension of machine learning methods in which domain experts provide additional knowledge by giving arguments for incorrectly classified cases. ABML typically extends methods that learn classification rules. The most advanced techniques for image classification tasks use convolutional neural networks (CNNs), which we are enhancing with ABML principles.
The main challenge is incorporating ABML principles into a deep learning model, as these models are often difficult to interpret. Moreover, domain knowledge for image classification is challenging to interpret. One suitable technique for explaining CNN model decisions is class activation mapping (CAM). Domain expert arguments can be captured using a computer mouse and then incorporated into the model. By comparing CAMs with formal arguments, domain knowledge can be integrated into the model.
Labels for constructing arguments were captured using a developed program and a computer mouse. The analysis of the model's performance was conducted using synthetic images of geometric shapes. We showed an improvement in performance for the ABML model for image classification even with different kinds of arguments.
|
