Loss Valleys, Uncertainty, and Generalization in Deep Learning/How do we build neural networks we can trust?