The features they use are actually fairly high level and thought out. I wonder if this could work in a boosting framework where you throw in thousands of really "dumb" features and let it combine them in a meaningful way, such that it discovers the high level concepts they are trying to model by itself.