On Deep Ensemble Learning from a Function Approximation Perspective