ARCHIVES

Handwritten digit recognition is essential in applications like postal sorting, cheque verification, and document digitization. While convolutional neural networks (CNNs) achieve high accuracy in image classification, their complexity and resource demands limit their use in low-power or transparency-focused environments. This study compares a classical logistic regression model with a CNN for digit classification on the MNIST dataset using PyTorch. The MNIST dataset consists of 60,000 training and 10,000 test grayscale images (28×28 pixels). Logistic regression flattens each image into a 784-dimensional vector and applies a linear classifier optimized with cross entropy loss and stochastic gradient descent. The CNN, on the other hand, uses convolutional layers to automatically extract spatial features for classification. Experimental results show logistic regression achieves over 90% accuracy, demonstrating efficiency and interpretability. Meanwhile, the CNN surpasses 98% accuracy, highlighting the advantage of deep learning for complex feature extraction. This comparison underscores the trade-off between model simplicity and performance. Logistic regression serves as a strong baseline or practical solution when resources and interpretability are priorities, while CNNs offer superior accuracy for more demanding tasks.