Quantifying Epistemic Uncertainty in Deep Learning