Reconciling Modern Deep Learning with Traditional Optimization Analyses: The Intrinsic Learning Rate