Clustering Inductive Biases with Unrolled Networks