Explainable AI: Improving the clarity of attention maps by differentiable binarization

Visual explanation enables humans to understand the decision-making process of neural networks. The attention branch network architecture is a promising approach for explainable neural networks. It generates attention maps highlighting the regions in the image relevant for classification. In this thesis, investigations are being carried out to what extent the significance of attention map values can be further increased. As an approach, multiple loss functions were applied over the attention values and evaluated the resulting attention maps. The introduction of the loss function over the attention maps helped in obtaining concise and differentiable binarized attention maps while keeping the model performance similar. This improves the clarity in the attention maps which in turn improved the explainability of the model.