The premise of map-reduce seems to basically be, "if you're asking questions of a set of objects then you can divide that set by the number of cores you have and ask questions in parallel".

Doesn't that seem intuitively obvious? There sure are a whole lot of papers being spawned over this obvious thing. I'll go finish the paper now and see if it's simply about converting machine learning algorithms into "ask questions of a set of objects" form.

Many machine learning architectures do not immediately suggest scalable training algorithm implementations.

This paper was accepted to NIPS, so apparently some bigwigs thought it was important.

Ah. Neat hack then!