2% is HUGE at this point, at least on the datasets that I am familiar with - ImageNet, MS-Coco, PascalVOC etc. And at this point, any modifications or strategies that gets you the 2% improvement is noteworthy, and I know that people in my team are looking forwards to techniques that will give us these improvements.
Hell, on MNIST 0.14% is huge. Geff Hinton created an entire new architecture to get 0.25% error, which is far better than the baseline 0.39% error [1].
To be fair, he did reduce the error by 35%.