Transfer learning in classifying handwritten Hebrew-script letters

Abstract

Classification is a fundamental task used in machine learning where the goal is to predict the category or class of given data points based on certain features. It is widely used across various fields especially in image recognition. One of the most classic and pedagogic examples is classifying handwritten letters or numbers. Several robust datasets have been established in various languages, particularly in English, enabling individuals to develop highly accurate letter classification models.¶ Interestingly, there appears to be no readily accessible dataset for the classification of handwritten Hebrew script. Some advancements in this area have emerged from Ben-Gurion University of the Negev, where a team developed a modest dataset of Hebrew script letters for the task of classification called the Hebrew Handwritten Dataset1. However, due to the small dataset, no model has been able to reach the benchmark results other datasets have reached in the past