CoordConv: Addressing an Intriguing Failing of Convolutional Neural Networks
At this meetup, we'll be having a look at a relatively new addition to the deep-learning toolkit: CoordConv. The approach works by adding the coordinates of pixels and feature maps to the network layers.
Introduced last year by Uber Engineering, it's a neat idea which they demonstrate improvements across a range of tasks:
You can introduce this to almost any layer in your network and if your task involves localization or can benefit from it, CoordConv may just help you!
R. Liu, J. Lehman, P. Molino, F. P. Such, E. Frank, A. Sergeev, and J. Yosinski, “An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution”
In the original paper, they create a few dummy tasks to better showcase the precise advantages CoordConv introduces. You can find pictures of these tasks in the paper and some additional animations in the blog post below. - Uber Engineering blog post: https://eng.uber.com/coordconv/
Video explanation (this is also embedded in the blog post above): https://youtu.be/8yFQc6elePA There was quite an interesting reaction from the community (see Reddit link). Particularly amusing was a brutal takedown in a blog by Filip Piekniewski.
See you there!
A note about the Journal Club format:
The sessions usually start with a 5-10 minute introduction to the paper by the topic volunteer, followed by splitting into smaller groups to discuss the paper and other materials. We finish the session by coming together for about 15 minutes to discuss what we have learned as a group and ask questions around the room.
There is no speaker at Journal Club. One of the community has volunteered their time to suggest the topic and start the session, but most of the discussion comes from within the groups.
You will get more benefit from the session if you read the paper or other materials in advance. We try to provide (where we can find them) accompanying blog posts, relevant code and other summaries of the topic to serve as entry points.
If you don't have time to do much preparation, please come anyway. You will probably have something to contribute, and even if you just end up following the other discussions, you can still learn a lot.
It’s OK just to read the blog post or watch the video :)
We don’t have spare copies of the paper during the session, so please print out your own if you want a hard copy for discussion. For digital copies, you are welcome to use your laptops/tablets/phones during the session.