Neural Network Implementation for Handwritten Digit Recognition without Using Any Libraries in Java

This Java project implements a neural network from scratch to recognize handwritten digits from 0 to 9, trained on the MNIST dataset. The network architecture, training process, and future enhancements are detailed below.

Project Overview

Network Architecture

• Input Layer: 784 neurons (28x28 pixel images)
• Hidden Layer: 10 neurons, activated by the Rectified Linear Unit (ReLU) function
• Output Layer: 10 neurons, activated by the softmax function

Training Algorithm

The neural network is trained using the Stochastic Gradient Descent (SGD) Backpropagation algorithm.

Future Enhancements

• Graphical User Interface (GUI):
  • Implemented using JavaFX.
  • Allows users to draw a digit on the screen, which the neural network will then predict.
• Utils_Multi Class Update:
  • The current implementation runs out of memory.
  • An updated version will be released soon.

Additional Features

• Test Class:
  • Evaluates the neural network on the MNIST test dataset.
  • Calculates and reports the accuracy of the neural network.

Technical Details

• Programming Language: Java 22
• Build Tool: Maven
• OS: Arch Linux x86_64