Likelihood-ratio-based confidence intervals for neural networks