How do I learn how to make novel algorithms?

I noticed that some jobs desire the candidate to be able to produce novel machine learning algorithms.

Although I have some grasp of the theory behind Machine Learning as I've taken courses on it, I still wouldn't have a clue as how to design algorithms. I mean I understand Max. Likelihood, Naive Bayes, the derivation of the EM algorithm via Jensen's inequality, inference in belief networks etc.

So I have more understanding of the theory than some who just use a bag of libraries to crunch data. But it seems a huge chasm from going from some mathematical understanding to formulating your own methods!?