Friday, April 04, 2014

Deep ConvNets; "Astounding" baseline for vision

Last night, Pierre Sermanet gave a talk on OverFeat and ConvNets. The meetup was triggered by one of Pierre's post on Google+ where he mentioned an ArXiv preprint where some folks at KTH [2] used features from OverFeat [1,3] and an SVM to get results they called "astounding" and said so in the title of the paper ( CNN Features off-the-shelf: an Astounding Baseline for Recognition ). The use of such wording is rare (If you recall similar words such as "stunning" or "quite striking" were used for a different 2011 paper in compressive sensing). So at the Paris Machine Learning Meetup group, we decided to have an impromptu specialist talk/meetup while Pierre was around. Gabriel Synnaeve got his lab to co-sponsor the talk that was eventually held at Normale Sup. Pierre made available the two most important slides of his talk here.

Mentionned in the talk besides OverFeat were:
More on this later and if anything it tells me the sensor designers ought to pay more and more attention to this type of breakthrough.

Thank you Pierre for the talk, and thanks to Gabriel for co-organizing the meetup. Other thanks go to Criteo, DoJoEvents and Alexandre for trying hard to fit our short fuse meeting with their hosting capabilities. About 60 people showed up with a two days notice !

I am not sure the Hangout on Air worked and the video from my GoPro has very little sound.

[1] OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks, Pierre Sermanet, David Eigen, Xiang Zhang, Michael Mathieu, Rob Fergus, Yann LeCun,

[2] CNN Features off-the-shelf: an Astounding Baseline for Recognition, Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, Stefan Carlsson,


JD said...

Hey Igor, thanks for all this great work putting the blog together.

I tried joining your Hangout a bunch of times within the half hour after it was supposed to start, but only saw a screen saying the Hangout would begin soon. Just letting you know in case it helps you get it right the next time.

Igor said...


I apologize for this, I screwed up there. It'll be better next time, I sweae.