From the early days of machine learning, background knowledge and data have been seen as two sides of the same coin. In more recent times, the use of knowledge has declined. One reason, I suggest, is that knowledge of sufficient specificity is often hard to find in many domains. The knowledge available is considered too general to be easily exploited. In this presentation, I discuss a particular system that successfully uses very abstract knowledge. This theory is known to only weakly approximate the dynamics of any practical system. Yet it is very useful is learning the multidimensional functions capturing these dynamics. I will discuss what it is about this approach that makes it effective. Through discussion with the audience I hope we can all come away with a clearer idea of when and how very abstract knowledge can be successfully employed in learning.