What is the bias-variance tradeoff in machine learning?