@amasad @tszzl In a high dimensional optimization space, on average you should expect to see few points that are hard to escape in all directioms and lots of saddle points. But the dimension on which you can escape your current saddle point is often hard to find!
@amasad@tszzl在高维优化空间中，平均而言，你应该看到很少几个在所有方向上都难以逃脱的点和很多鞍点。但你能逃离当前鞍点的维度通常很难找到！