A Modern Take on the Bias-Variance Tradeoff in Neural Networks