“Initializing a deep neural network by pre-training with a stack of Restricted Boltzmann Machines allows it to learn much faster and generalize better than initializing with random weights.”