nlib

A simple C deep learning library I wrote back in school for a project that I used for experiments with handwriting recognition.

The code hasn't been touched in 10 years and obviously shouldn't be used by anyone; this repository just exists so the code doesn't get lost.

Features

• A simple and easy-to-use interface; creation and training of networks in just 3 lines of code
• Creation of networks via xml-files
• Support for FullyConnected, Convolutional and Pooling layers
• GPU acceleration via cuDNN v5
• Training via backpropagation (although training of Convolutional and Pooling layers is currently only supported with CUDA)
• Momentum
• Committees of networks
• Activation functions: Sigmoid, Softmax, ReLu, leaky ReLu, Softplus
• Cost functions: Quadratic/MSE, Cross Entropy
• Pooling methods: Average and Maximum
• Regularization: L2
• An easy way to load and manage datasets such as the MNIST-database

Dependencies

• GSL (GNU Scientific Library)
• GLib
• libxml2
• CUDA 7.5 and cuDNN v5 if compiled with CUDA support

Planned features

• Training of Convolutional/Pooling layers without CUDA; better optimized Convolution-forward algorithm
• Dropout regularization
• Recurrent neural networks