random thoughts on this announcementI took Yann's ML class during the first semester of my program at NYU two years ago. It was a terrible experience, given that I hadn't worked with linear algebra or mathematical statistics since about 2008. I could barely write Python at the time, so my Lush code was some ungodly non-functional crap. That said, I'm very happy I took the class in 2011 despite hyperventilating my way through the final--it was only offered once more, and I suspect it will never be offered again given today's news.
Yann really knows his stuff. His convolutional nets (and their application to MNIST and image segment classification) represented a significant improvement in computer vision, and he demoed some incredible low-latency image segmentation stuff for us. He runs one of the best neural net labs in the world (up there with Hinton's and Bengio's), and he has some incredible students at NYU. I can see this shaping up as a delayed acquihire of sorts...he will not struggle to find excellent candidates.
There's a lot of discussion around the privacy implications, but I think everyone's rehashing old points. Facebook already hires excellent researchers and data scientists -- John Myles White and Sean Taylor just moved out west -- but they don't focus on images just yet, from what I know. This hiring represents an investment in image analysis on Facebook's behalf that matches what they put into unstructured textual data and graphical inference. If you've already stayed with Facebook through the "graph search" announcement, this shouldn't surprise you either.
As someone interested in this type of work, it's an exciting time to live in New York City. Finance and adtech have been here forever, but things have expanded. Many startups (Foursquare, Tumblr, Knewton, Etsy) have invested in applied statistics and machine learning, hiring excellent researchers and engineers. Columbia and NYU have announced data science initiatives in the last 6 months. Very smart people (and others, like me ;) ) are very active in the community here.
There are some obvious applications of Yann's work to Facebook's advertising goals:
> Identifying strong friendships through co-occurrence in photos
> Digit or character recognition applied to marketing in photos (shop signs, brands, etc)
> Image segment classification (e.g. beach, park, road) for use in predicting a photo's location (for those uploaded after the fact)
> All the Instagram photos. I mean seriously. They're committed to making ads seamless--why not use image segmentation + likes to identify photographic structures people are attracted to?
Can anyone think of others?