Variance-Covariance Regularization Improves Representation Learning