right now such a parallel track doesn't really exist, at least not on an even footing. as the article shows, there's still a heavy focus on beating accuracy be

right now such a parallel track doesn't really exist, at least not on an even footing. as the article shows, there's still a heavy focus on beating accuracy benchmarks, at huge computational cost, in terms of what actually gets accepted at major venues. it can be quite difficult to make the case for a method that has worse performance, but is cheaper. hard to judge without real data but I think such papers are much less likely to be accepted. there are also sometimes demands to replicate very expensive techniques as baselines, which can be onerous for groups with limited resources.