How to Escape Saddle Points Efficiently – The Berkeley Artificial Intelligence Research Blog