MNIST Digit Recognition Neural Network

🎯 Project Overview

A pure Java implementation of a neural network for recognizing handwritten digits from the MNIST dataset. This project was built from scratch without relying on any machine learning libraries to deepen understanding of neural network fundamentals.

🧠 Neural Network Architecture

The network consists of three layers:

• Input Layer (784 neurons): Processes flattened 28×28 pixel images
• Hidden Layer (128 neurons): Fully connected layer with ReLU activation
• Output Layer (10 neurons): Produces predictions using softmax activation

🔑 Key Features

• Pure Java Implementation: Built without external ML libraries
• MNIST Dataset Support: Handles the standard MNIST dataset format
• Xavier/Glorot Initialization: Smart weight initialization for better training
• Backpropagation: Implements complete gradient descent learning
• Real-time Training Metrics: Monitors accuracy during training
• Validation: Tests network performance every 5 epochs

📊 Technical Details

• Activation Functions:
  • Hidden Layer: ReLU (Rectified Linear Unit)
  • Output Layer: Softmax
• Learning Parameters:
  • Learning Rate: 0.01
  • Epochs: 10
  • Batch Size: 1 (Stochastic Gradient Descent)

🚀 Getting Started

Prerequisites

• Java JDK 21
• MNIST dataset files:
  • train-images.idx3-ubyte
  • train-labels.idx1-ubyte
  • t10k-images.idx3-ubyte
  • t10k-labels.idx1-ubyte

Dataset Structure

Place the MNIST dataset files in the dataset/ directory:

Running the Project

Run these two commands:

javac src/.java

java src/Main

📈 Performance

The network typically achieves:

• Training Accuracy: ~90-95%
• Test Accuracy: ~90%

Screenshot