Look beyond labels: Incorporating functional summary information in Bayesian neural networks