Introduces curriculum learning: training models on examples of increasing difficulty. Shows this acts as a continuation method for non-convex optimization, improving both convergence speed and final generalization.