Incorporating functional summary information in Bayesian neural networks using a Dirichlet process likelihood approach