Multimodal Deep Learning Library

This is the document of Multimodal Deep Learning Library, MDL, which is written in C++. It explains principles and implementations with details of Restricted Boltzmann Machine, Deep Neural Network, Deep Belief Network, Denoising Autoencoder, Deep Boltzmann Machine, Deep Canonical Correlation Analysis, and modal prediction model. MDL uses OpenCV 3.0.0, which is the only dependency of this library. Most of its implementation has been tested in Mac OS. It also provides interface for reading various data set such as MNIST, CIFAR, XRMB, and AVLetters. To read mat file, Matlab must be installed because it uses Matlab/c++ interface provided by Matlab. There are multiple model options provided. Different gradient descent methods, loss function, annealing methods, and activation functions are given. These options are easy to extend given the structure of MDL. So MDL could be used as a frame for testings in deep learning.