MNIST_C: Neural Network Digit Recognition

A lightweight C implementation of a neural network for handwritten digit recognition using the MNIST dataset. This project provides a simple yet effective command-line tool for training, testing, and applying a neural network model to recognize digits.

Features

• Feed-forward neural network with configurable parameters
• Training using backpropagation algorithm
• Testing against the MNIST test dataset
• Recognition of digits in custom image files
• Persistence capabilities to save and load trained models
• Black & white image conversion option

Requirements

• C compiler (gcc recommended)
• MNIST dataset files (not included)
• Standard C libraries

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/MNIST_C.git
   cd MNIST_C

2. Compile the source code:

   make

Usage

The program provides a command-line interface with the following commands:

USAGE:
  ./neural <command> [options]

COMMANDS:
  learn      Train the neural network with specified parameters
  test       Evaluate the network's performance on test data
  learnTest  Train the network and then test its performance
  apply      Recognize a digit in the specified image file

OPTIONS:
  learning_rate  Learning rate (default: 0.3) - Controls step size during training
  attempts       Number of training iterations (default: 2)
  blackNwhite    Use black and white conversion [0|1] (default: 1)
  path           Path to image file for digit recognition

Examples

# Train with learning rate 0.1, 5 iterations, B&W mode
./neural learn 0.1 5 1

# Test pre-trained network using B&W mode
./neural test 1

# Train and test with specified parameters
./neural learnTest 0.2 3 1

# Recognize digit in the specified image
./neural apply digit.png 1

Architecture

The neural network consists of:

• Input Layer: 784 neurons (representing a 28x28 pixel image)
• Hidden Layer: 300 neurons
• Output Layer: 10 neurons (one for each digit 0-9)

The network uses the sigmoid activation function and is trained using backpropagation with a configurable learning rate.

File Structure

• Neural.c/h: Main neural network implementation
• MathTools.h: Mathematical functions for neural network operations
• Persist.h: Functions for saving and loading trained models
• DataLoad.h: Functions for loading MNIST data and processing images

Performance

On the MNIST test dataset, the network can achieve accuracy rates that vary based on training parameters. With optimal settings, it can reach significant accuracy while maintaining computational efficiency.