Top interpretable neural network for handwriting identification

Machine learning (ML) has become one of the most promising tools in forensics, despite its dominant method of artificial neural networks (ANNs) suffering from the black-box problem. While forensic methodology demands explainability and evaluativity, neural networks are unexplainable, hence almost unfalsifiable. Our study was conducted to mitigate this problem in a case-like context, by creating a novel top interpretable neural network (TINN) for identification of the authors of handwritten documents. The idea of top interpretability assumes that it is irrelevant how the handwriting features are extracted from documents, as long as they are semantically sensible and the sole determinants of identification. The model was tasked with supervised extraction of handwriting characteristics and subsequent identification of the writers on that basis. The interpretable model not only outperformed all comparative models in terms of author identification, but also underperformed in terms of features extraction (achieving satisfactory results nonetheless). Visualizations of features-extracted by the model to perform its tasks-suggest that it considers rational and semantically sensible features of handwriting, but we are unable to determine whether it learned the exact features of handwriting we desired. The approach of top interpretability proved to be effective in terms of accuracy and interpretability. Furthermore, if we were to judge features extracted by such a network as unconvincing, then our approach is a highly efficient method of falsification. Lastly, the success of this study-performed on a small-scale identificational problem-suggests that a similar approach could yield better results on a large-scale identificational or verificational problem.