Understanding the Loss Surface of Neural Networks for Binary Classification