This is an attempt to program a neural net from scratch in C, to test my understanding of deep learning fundamentals.

This repo contains a .c implementation of a flexible, fully connected feed-forward neural net, whose functions are accessible via network.h.
Layer dimensions, count, and other hyperparameters are adaptable by changing the macros.

Im using it to classify 28x28 images of numbers. The network has size 784x300x10 and is trained on MNIST.
Im using softmax + cross-entropy for the output layer, LeakyReLU for the hidden layer, and basic gradient descent for backpropagation.

TODO: refactoring and adding saving and loading funcs

Future Plans

I plan on further optimizing by using Adam and mini-batches.

Usage

To build the project, simply run:

make

to train MNIST

./main

License

MIT License