Nuclear Norm Maximization Based Curiosity-Driven Learning