XOR Neural Network in C

This project implements a small feedforward neural network in C that learns the XOR function using the backpropagation algorithm. The program demonstrates how basic neural networks can be built and trained from scratch without relying on external libraries.

Features

• Custom Implementation: All aspects of the neural network, including forward propagation, backpropagation, and weight updates, are implemented in pure C.
• Adjustable Parameters:
  • Learning Rate: Configurable via the learningRate variable.
  • Epoch Count: Adjustable through the epochCount variable for controlling the number of training iterations.
  • Network Architecture:
    • 2 input nodes
    • 2 hidden layer nodes
    • 1 output node
• Shuffling Training Data: Uses the Fisher-Yates algorithm to randomize the order of training samples in each epoch.
• Sigmoid Activation Function: Used for both hidden and output layers, with its derivative employed in backpropagation.

How It Works

• Initialization:
  • Weights and biases are randomly initialized to small values in the range [0, 1).
• Forward Pass:
  • Inputs are propagated through the hidden and output layers using the sigmoid activation function.
• Backward Pass:
  • Errors are computed at the output and propagated back to adjust weights and biases using the gradient of the sigmoid function.
• Training:
  • The network is trained on the XOR dataset:

    Input: 0, 0 -> Output: 0
    Input: 0, 1 -> Output: 1
    Input: 1, 0 -> Output: 1
    Input: 1, 1 -> Output: 0
    

  • After sufficient epochs, the network learns to predict the XOR function accurately.

Adjustable Parameters

• Learning Rate: Modify learningRate to control the step size of weight updates.
• Epoch Count: Adjust epochCount to train the network for a specific number of iterations.

Compilation and Execution

1. Clone the repository.
2. Compile the program using gcc:

   gcc -o xor_nn nn.c -lm

The -lm flag links the math library for functions like exp.

3. Run the program:

   ./xor_nn
   

Output

The program prints the network's predictions for each input alongside the expected outputs during training. Over multiple epochs, the predictions converge to match the expected XOR outputs.

Example output:

Input: 0 0 Output: 0.003 Expected Output: 0
Input: 0 1 Output: 0.978 Expected Output: 1
Input: 1 0 Output: 0.982 Expected Output: 1
Input: 1 1 Output: 0.005 Expected Output: 0

Future Improvements

• Add support for more complex activation functions.
• Implement momentum-based gradient descent for improved training efficiency.
• Generalize the network to support arbitrary architectures.