I've been quite disappointed to see how many papers from "cutting edge" research groups are chasing small improvements in well-known benchmarks by finding new t

I've been quite disappointed to see how many papers from "cutting edge" research groups are chasing small improvements in well-known benchmarks by finding new techniques that happen to work, while there is a lot less effort put into finding out why. I guess Geoff Hinton's explanation about what gets published today explains it.