I find it interesting that now Ockham's razor moved from philosophy to applied statistics. That is, Bayesian statistics can quantify (a more complicated model which fits just as well has, or only slightly better, has lower likelihood) and in machine learning we use it in practice (to avoid overfitting, as too complex models may fit well to the training data, but be suboptimal for generalization).

See also BIC (Bayesian Information Criterium) for selecting models.