Approximate computation and implicit regularization for very large-scale data analysis