NeuralNets

Small neural network library written in C from scratch.

Features

• Mini-batch gradient descent
• Mean squared error
• Cross-entropy cost
• Sigmoid activation
• L2 regularization (weight decay)
• Momentum

Usage

Run tests

make test
./build/test

Train xor

make xor
./build/xor

Train MNIST

• Download the dataset, convert to csv and save to resources folder

make mnist
./build/mnist

MNIST Results

~97% validation accuracy after around 10 epochs

• sigmoid activation and cross entropy cost
• single hidden layer with 100 hidden neurons
• learning rate: 0.1
• batch size: 10
• momentum 0.9
• L2 lambda: 0.09

See examples/mnist.c for details. For MNIST digits classification you need to have the dataset converted to csv format.