Deep Torah Learning: A Deep Learning Approach to Identifying Anomalous Letters in Torah Scrolls

Abstract

Otiyot meshunot, anomalous letters in Torah scrolls, can be useful in studying the historical evolution of Torah scrolls, as well as determining the chronological and geographical origins of a particular scroll. Currently, otiyot meshunot are discovered by researchers manually examining each letter in a scroll. In order to improve the efficiency of this process, a deep learning model was built to distinguish anomalous Torah letters from normal ones. The model, trained on images of Torah scrolls, used a neural network to accomplish optical character recognition of typical Torah letters and anomaly detection of unique letters. The model succeeded in identifying numerous otiyot meshunot, and serves as a prototype for a potential product to be used in the realm of Torah research.