This program is a very basic implementation of a feedforward neural network in C. It is more or less a proof of concept, as the plan is to port the NN to a functional language in the future. It uses a half-baked matrix library that was quickly thrown together as the basis for the operations performed. It uses a sigmoid activation function for the neuron's activation, and gradient descent backpropagation to train the network.
git clone [email protected]:whazell/C-Machine-Learning.git
cd C-Machine-Learning
mkdir build && cd build
cmake ../
make
This program comes with 2 examples, sin_test and xor_test. Their usage is outlined below.
./build/bin/examples/xor_test
./build/bin/examples/sin_test
Currently there are two basic examples for this neural network. The first is xor_test, which trains the network to predict the xor function. The second is the sin_test which predicts values from the sin() function.
This uses a basic 2 input, 8 hidden, 1 output node structure. It can accuratly predict the function after a few hundred epochs. It loads the data from the file tests/xor.csv.
The is used to predict the sin() function, by feeding in normalized results from the sin() function (tests/sin-gen.c builds the data). It uses a single input node, 2 hidden layers of 5 nodes, and a single output layer. The data has to be normalized as the sigmoid function only produces values on the range (0,1) while the sin function produces on the range (-1, 1). By applying the transformation 0.5 * (sin(x) + 1), we can now train the neural network on this function, and undo the transformation after the value is predicted (ie (2 * result) - 1). This is very successful at predicting the values of sin().
This project is built with the munit test suite. The tests are built automatically when cmake is ran. They build the tests into the directory build/test/. Currently ther are the test suites:
matrix_test
Run with: ./build/test/[test_suite_name]
All public interface functions are preceeded by 'gn' to ensure no namespace conflicts. Currently, the API consists of very basic functions and is not stable at all.