Reconciling modern machine learning and the bias-variance trade-off