Computer Vision News - June 2022

16 Outstanding Paper Award On behalf of the community, we have to ask, does he have any advice about what we should do when working on something and a solution is just not coming? “ If I had a sure-fire answer to that, I would be halfway to the Nobel Prize! ” he laughs. “ In my experience, when things aren’t working out, simplify. Even if it doesn’t turn out that the simplest possible algorithm is the right one, it’s much easier to see what’s goingwrong. I apply a kind of Occam’s razor principle that the simpler the algorithm is, the more widely applicable it will be . Rich Sutton’s Bitter Lessons are all about this. Of course, you might have to increase complexity again at some point. Then it’s important to know where you can afford to have that complexity so that it still works. ” We ask Sebastian what he thinks convinced the jury that his paper deserved an award, andhepoints to three things. The first is that the idea is so simple and general, making it of interest to the community because it could be taken off in many different directions. Secondly, the striking results the team achieved on Atari. Thirdly, the way the paper was written seemed to appeal to people in terms of the mix of inspiration and theoretical and empirical insight. “ I did everything I could to make sure the Atari results were a fair comparison and ablated all our choices, ” he adds. “ It stands out that you can double your performance by tuningthesehyperparameters online if you do it correctly. guarantees for this, but it looks nice on paper, and pretty quickly, we got very strong results, ” he recalls. “ The empirical stuff is where you usually spend time making sure things work. But here, once we had the right algorithm and applied it, it worked straight out of the box. The hardest part is spending enough time on the abstract problem to have the insight to propose the right algorithm. ”