A Differentiable Rank-Based Objective For Better Feature Learning