IncorporatingInterpretableOutputConstraints inBayesianNeuralNetworks