neural-network

A lightweight, user-friendly Neural Network written from scratch without external libraries. It includes forward- and backpropagation as well as gradient descent.

Features

• 2D Matrix Library (Matrix2D): Custom implementation supporting Matrix Multiplication, Addition, Subtraction, Transposition, and more which are required for Neural Networks.
• Layers: Implements a base Layer class as well as two subclasses (LinearLayer and ActivationLayer).
• Activation Functions: Built-in implementations of ReLU, Tanh, and Sigmoid alongside their derivatives.
• Loss Functions: MSE (Mean Squared Error) and CEL (Cross-Entropy Loss) also come built-in.
• Flexible Network Architecture: Features a Sequential class similar to that of Pytorch, where you can customize the layout of your Neural Network easily.

Prerequisites

• C++ Compiler supporting C++23
• GoogleTest (only required for running the test suite)

Running the Main Application

You can quickly compile and run the main application by using the included Makefile.

make
./main.exe

Running the Tests

The same applies for compiling and running the tests.

make tests
./tests.exe

Usage

#include "matrix.h"
#include "nn.h"

int main() {
    Matrix2D inputs;
    Matrix2D outputs;
    
    inputs.addRow({0.1, 0.2});
    outputs.addRow({0.2});

    // Define Model Architecture
    Sequential model(
        LinearLayer(2, 4), 
        ActivationLayer(
            NeuralNet::ApplyTanh, 
            NeuralNet::DTanh
        ),
        LinearLayer(4, 1)
    );

    // Train the model (inputs, targets, epochs, learning rate)
    model.train(inputs, outputs, 1000, 0.1);

    return 0;
}