Using Sum-Product Networks to Assess Uncertainty in Deep Active Learning