As a counterpoint, I seems to me that the ML field is too heavy on theory. A lot of the theoretical papers are proposed with weak or no experimental tests. There are very few people doing thorough experimental work. What I have in mind is something like Chen and Goodman's comparison of different language models - this is a great paper and I'd love to see more like it.
The obvious example of this is boosting where the theory lagged far behind the experiment - everyone wanted to invent a theory to explain why boosting was so effective.
hey industrialists out there - publish!
(I mean, sure, you can try to fire-and-forget, but for the most part you'll either fail with no clue why, or fail to get very good results. Decision trees might be a moderate exception, although even then you really ought to understand how they are constructed and the implications of various choices you can make.)