Frankly, I have mixed opinions about this blog post. Good for discussing types of papers, and that for the toy you can write convolutions by hand (which, IMHO,

Frankly, I have mixed opinions about this blog post. Good for discussing types of papers, and that for the toy you can write convolutions by hand (which, IMHO, is by not means any argument against CoordConv!). I adore toy problem they (author of the paper) picked, and if anything, it is an argument for their choice of the toy problem (unsolvable by typical conv, trivial when added x and y channels). In science it is crucial to make many failed approaches, not only approaches of things that we are sure they work. So yes, it's good that they burnt 100 GPUs on a problem that didn't work. And in fact it is much better standard than most deep learning papers I read, when they focus mostly or only on problems in which architecture is better. Plus, it works for object detection, so it's not a "MNIST-only trick".