Unsupervised Pre-Training Speeds up the Search for Good Features : An Analysis of a Simplified Model of Neural Network Learning