"Hi Serge"

Back in early 2012, we submitted a paper to CVPR 2012 describing the first version of our convolutional network-based semantic segmentation system (or scene labelling system).

The paper improved the state of the art for scene labelling on three datasets and was shown to be faster than the closest competitor by two orders of magnitude.

Nevertheless, the paper received very negative reviews ("definitely reject", "weakly reject", "berderline"). The reviews appeared to us as misinformed and biased. It was clear that the reviewers either had a negative bias against neural nets, feature learning, or deep learning, or  had no idea what a convolutional net was, or both.

This followed a long string of rejected paper about convolutional nets. Given this negative bias, I decided to no longer send convnet-based papers to CVPR, because it's was a complete waste of time, energy, and good will.

Rather than sending a rebuttal, we decided to pull the paper and submit it to ICML, where we thought the reviewers might be more positively disposed.

When we withdrew the paper, I sent a letter to the program chair +Serge Belongie, to explain our reasons. Serge decided to post an anonymized version of my withdrawal letter, together with the reviews: https://docs.google.com/document/d/1-KxI1oZl9A0DMewZBJArUPPKUrVLYr4FrDey5d_3QR0/edit
As I said in the letter, I did not blame Serge (I was chair of CVPR 2006. I know how it is). It was a cultural/sociological/structural problem that nobody should be blamed for. I wasn't bitter, or angry about the CV community. I was just merely trying to minimize our waste of time and mental energy.

Despite my pledge to avoid CV conferences, I wasn't going to stop my students from submitting if they really wanted to. We did get a couple of convnet-based papers published in ECCV12 and CVPR13, as well as other papers that weren't about convnets (e.g. fast sparse coding).

Our withdrawn paper was eventually accepted to ICML12: http://yann.lecun.com/exdb/publis/index.html#farabet-icml-12
A video of +Clement Farabet's ICML talk is here: http://techtalks.tv
/talks/57300/

The work was extended and turned into an IEEE T.PAMI paper for the upcoming special issue on deep learning: http://yann.lecun.com/exdb/publis/index.html#farabet-pami-13

This line of work was further extended by +Camille Couprie to RGB-Depth images and video:
http://yann.lecun.com/exdb/publis/index.html#couprie-iclr-13
http://yann.lecun.com/exdb/publis/index.html#couprie-icip-13

Since the ImageNet competition was smashed by the +Alex Krizhevsky/+Ilya Sutskever/+Geoffrey Hinton  convnet in October 2012, the attitude of parts of the computer vision community towards convnets has been evolving.

In fact, it could be argued that convnets and deep learning are all the rage now, if I judge by how many people showed up at my invited talk at the CVPR Scene Understanding Workshop Sunday (and also by the fact that I actually had an invited talk).

Because of this, I am happy to announce that I am no-longer planning to avoid submitting deep-learning papers to computer vision conferences.
Shared publiclyView activity