The Pitfalls of Simplicity Bias in Neural Networks