The Role of Proportionality in Five Hidden Layers Back Propagation Neural Network

In this study, we analyze the impact of neuron proportionality across hidden layers on the accuracy of neural network outputs. Our findings suggest that optimal neuron distribution, combined with an appropriate number of hidden layers, can significantly enhance accuracy. While typically two hidden layers suffice for training, scenarios demanding high accuracy, such as multiscript numeral recognition, necessitate more layers. This is particularly crucial when dealing with numerals of similar shapes but distinct values across different scripts. For instance, the shape '0' in Farsi resembles the numeral 'Five', whereas in Hindi, English, and other scripts, it represents 'Zero'. Such challenges are prevalent when numbers or postal codes are written in multiple scripts, such as Hindi and Farsi. We address these issues, emphasizing accuracy as a primary criterion.