This study addresses the ongoing pursuit of  achieving optimal performance in digit recognition tasks, 
focusing on the widely studied MNIST dataset. Our motivation  stems from the challenge of accurately classifying the 
remaining 1% of images, despite the relatively high 99%  accuracy achieved by existing models. In this work, we present 
a simplified approach to convolutional neural network (CNN)  architecture, aiming to streamline model complexity while 
maintaining or even enhancing performance. Unlike previous  approaches, our methodology involves utilizing only two CNN  layers with fewer filters, resulting in a reduction in model  parameters and learning time. Through rigorous 
experimentation and evaluation, we demonstrate that our  streamlined CNN architecture yields competitive results. Our 
findings underscore the importance of exploring alternative  model architectures and optimization techniques to achieve  state-of-the-art performance in digit recognition tasks.