Understanding neural networks through sparse circuits