Tag: On the power and limitations of random features for understanding neural networks