Transparency Techniques for Neural Networks trained on Writer Identification and Writer Verification
This work addresses the need for interpretability in AI systems for forensic document analysis, but it is incremental as it applies existing transparency techniques to a new domain.
The paper tackled the problem of improving transparency in neural networks for Writer Identification and Writer Verification by applying two saliency map techniques, finding that pixel-level maps outperformed point-specific maps and are suitable for supporting forensic experts.
Neural Networks are the state of the art for many tasks in the computer vision domain, including Writer Identification (WI) and Writer Verification (WV). The transparency of these "black box" systems is important for improvements of performance and reliability. For this work, two transparency techniques are applied to neural networks trained on WI and WV for the first time in this domain. The first technique provides pixel-level saliency maps, while the point-specific saliency maps of the second technique provide information on similarities between two images. The transparency techniques are evaluated using deletion and insertion score metrics. The goal is to support forensic experts with information on similarities in handwritten text and to explore the characteristics selected by a neural network for the identification process. For the qualitative evaluation, the highlights of the maps are compared to the areas forensic experts consider during the identification process. The evaluation results show that the pixel-wise saliency maps outperform the point-specific saliency maps and are suitable for the support of forensic experts.