Provable Bounds for Learning Some Deep Representations