Stochastic Gradient Descent Neural Network in Pure C

A lightweight, efficient neural network implementation using stochastic gradient descent (SGD), written entirely in pure C. This project focuses on performance, simplicity, and a low-level understanding of neural network mechanics without the overhead of external libraries.

🚀 Features

• Fully connected feedforward neural network architecture
• Stochastic Gradient Descent (SGD) for training
• Dynamic network topology (adjustable layers and neurons)
• No dependencies—written in pure C

🧠 How It Works

The core of this neural network relies on:

1. Forward Propagation: Computes outputs layer-by-layer using weighted sums and activation functions.
2. Backpropagation: Calculates gradients for each weight to minimize error.
3. Stochastic Gradient Descent: Updates weights using gradients from random subsets (mini-batches) of data.

📦 Installation

git clone https://github.com/yourusername/sgd-neural-network-c.git
cd sgd-neural-network-c
make

📝 Usage

Training the Network

#include "include/math.h"
#include "include/structure.h"
#include "include/neural_network.h"
//include additonal header for handling strings and io
int main(){
	//linear input matrix of size 4 x 2
	Matrix * inputs = initMatrix(4,2);
	float inputsf[8] ={1,0,	1,1,	 0,1,	0,0};
	memcpy(inputs->matrix, inputsf,sizeof(float)*8);
	
	//linear input matrix of size 4 x 1
	Matrix * outputs = initMatrix(4,1);
	float outputf[4] = {1,	0,	 1,	0};
	memcpy(outputs->matrix, outputf,sizeof(float)*4);

	//layers Vector {2,2,1} 2 neurons at the input 2 inputs at the hidden layer 1 output
	Vector * layersV = initVector(3);
	float layers[3] = {2,2,1};
	memcpy(layersV->vector, layers,sizeof(float)*3);
	
	int layers_size=3;
	int learning_rate=1;
	neural_network nn = {
		.layers_size=layers_size,
		.layer_size = layersV,
		.learning_rate = learning_rate,
		.input =inputs,
		.output =outputs

	};
	construct_network(&nn);
	train_network(&nn,2000);
	nn.test_mode = 1;
	test_network(&nn);

}

Configuration Options

• Learning Rate: Adjustable to control the step size during optimization.
• Layers and neurons: Adjustable Layers and Neurons size.
• Epochs: Number of iterations over the training dataset.
• Activation Functions: Only Sigmoid avaliable now.

⚙️ Build & Run

make              # Compiles the program
make run  # Runs the neural network

📊 Example Dataset

Sample datasets for classification and regression tasks are provided in main.c(XOR) predicting.

🤝 Contributing

Contributions are welcome! Feel free to fork the repository, open issues, or submit pull requests.

📄 License

This project is licensed under the MIT License.

---

Note: This project is intended for educational purposes, demonstrating the fundamental principles of neural networks and SGD in C.