Feature Learning in L_2 -regularized DNNs: Attraction/Repulsion and Sparsity