Bayesian Neural Networks Avoid Encoding Complex and Perturbation-Sensitive Concepts