Comparative Study of Convolutional and Feedforward Neural Networks for Optical Character Recognition
Vishal Paul · 2024
This paper explores the performance of convolutional neural networks (CNNs) and fully-connected feedforward neural networks on optical character recognition tasks using the MNIST dataset. The study examines model accuracy and efficiency, highlighting the superior accuracy of CNNs, albeit at the cost of longer training and testing times. Modifications such as input fuzzing and adaptive learning rates are introduced to improve robustness and efficiency. While fuzzing enhances model resilience to noisy data, adaptive learning shows limited effectiveness due to underfitting in some cases. The paper proposes a hybrid model combining both networks to optimize classification speed and accuracy.