An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution

For any problem involving pixels or spatial representations, common intuition holds that convolutional neural networks may be appropriate. In this paper the authors show a striking counterexample to this intuition via the seemingly trivial coordinate transform problem, which simply requires learning a mapping between coordinates in (x,y) Cartesian space and one-hot pixel space. Their solution to this puzzling phenomenon is show to improve GAN, R-CNN and RL architectures.


Want to receive more content like this in your inbox?