Interesting read. The highest risk I foresee is to learn correlations, which is what can be done already with off-the-shelf machine learning and variable selection techniques.
However I consider the method sound and worth trying.

Keep us posted!