> Nobody evaluates a model based simply on how well it fits the data and its number of parameters

This is standard practice in deep learning...