Mathematical Foundations of Artificial Neural Networks and Their Application in Handwritten Digit Recognition

This study develops the mathematical foundations of artificial neural networks and demonstrates their application in handwritten digit recognition, using the MNIST benchmark dataset as a training and evaluation corpus within the context of advancing computational mathematics education in Nigerian universities. The proliferation of machine learning applications in technology, agriculture, healthcare, and finance in Nigeria has created an urgent demand for graduates with foundational competency in the mathematical structures underlying these systems, yet dedicated treatment of neural network mathematics within undergraduate mathematics curricula remains rare. The study presents a rigorous treatment of the perceptron model, multi-layer feedforward architectures, activation functions, loss functions, and backpropagation as a gradient descent algorithm operating on the composition of differentiable functions. The chain rule of calculus and linear algebra of weight matrices and bias vectors are developed in detail. A two-hidden-layer neural network is implemented from first principles in Python without high-level machine learning libraries, and trained on a subset of the MNIST dataset. The network achieves a test set classification accuracy of 97.8 percent after hyperparameter tuning, demonstrating the effectiveness of the mathematical model. The study concludes by discussing the relevance of these mathematical foundations for Nigerian students aspiring to work in data science, financial technology, and agricultural technology sectors. Keywords: artificial neural networks, backpropagation, handwritten digit recognition, machine learning mathematics, MNIST