Learning sparse features can lead to overfitting in neural networks

Leonardo Petrini, Francesco Cagnetta, Eric Vanden-Eijnden, Matthieu Wyart