Visual Steering for Deep Neural Networks Using Explainable Artificial Intelligence

Abstract

Despite their remarkable revolution in artificial intelligence and promising performance, deep neural networks are susceptible to biases due to imbalanced training data or the way that training data is represented as features for the model; thus, they need a high level of transparency in critical fields, especially in the medical sector. Several explainability methods have been developed, aiming at enhancing the transparency of the model and increasing its trustworthiness. One of these methods is Grad-CAM, which can highlight the significant parts of an input that most influence the model’s prediction and detect biases in the dataset. In this paper, a method is provided for integrating human knowledge with the deep neural network to improve its performance, reduce the biases that arise, and leverage human experience. This can be done via the attention maps generated by Grad-CAM, which are then edited by domain experts. The proposed method enables the edited attention maps to contribute to the training of the model and updating its parameters, leading to better generalisation and fewer inductive biases. This approach is applied to several image classification datasets: Imagenette2, BUSI, and skin cancer ISIC. Experimental results show that the generated attention maps highly correspond to the edited ones while having high classification performance.