Blind men come to agreement on position, orientation of elephant

January 4, 2018 8:44 AM

"This is an amazingly good video [by Aurélien Geron]. I wish I could explain capsules that well."- Geoffrey Hinton.
posted by a snickering nuthatch (11 comments total)

Interesting. Reminds me a lot of the Hough transform, only probably much more robust. Everything old is new again!
posted by simra at 9:16 AM on January 4, 2018 [2 favorites]

Yes! There are a lot of connections there.
posted by a snickering nuthatch at 9:36 AM on January 4, 2018

Some indication of what this post is about in the text, title, or linked URLs would be appreciated. Only the first tag gives me any clue that this isn't about, say, pills.
posted by one for the books at 9:37 AM on January 4, 2018 [19 favorites]

Apologies for the mystery meat. I should have been more explicit.
posted by a snickering nuthatch at 9:41 AM on January 4, 2018

Seconding One for the book's request for a bit more detail in the post. (Particularly for a video link, since it's harder to scan than an article.)
posted by dendritejungle at 9:41 AM on January 4, 2018 [2 favorites]

Since editing the post is not possible, I am guessing you are hoping that I will post more detail in a comment.

The post is about machine learning techniques. Recently one family of these techniques (deep convolutional networks) has enjoyed a lot of success. But Geoff Hinton, who himself has contributed to the development of those techniques, has been making noises about some problems with them. He finally put his money where his mouth was recently by releasing a paper (the "capsules" link) that provided an alternative method.

Now, Hinton is a big deal in ML circles and everyone knew that this paper could be important. But many people found the paper pretty hard to digest on its own. So various people have taken a whack at explaining what capsule networks are all about, but far and away the best (as judged by Hinton himself) is the first link in this post.
posted by a snickering nuthatch at 9:49 AM on January 4, 2018 [9 favorites]

I was hoping this was a video of blind men feeling an elephant and discussing it.
posted by matrixclown at 10:00 AM on January 4, 2018 [20 favorites]

I found the empirical results in capsule paper pretty unimpressive, TBH
posted by hleehowon at 10:18 AM on January 4, 2018 [1 favorite]

Thanks Jpfed, appreciate it!
posted by dendritejungle at 11:01 AM on January 4, 2018

damn, need to change my name to Aurélien, stat
posted by supermedusa at 11:03 AM on January 4, 2018

Geron is also the author of a fantastic introductory book, Hands-On Machine Learning with Scikit-Learn and TensorFlow Concepts, Tools, and Techniques to Build Intelligent Systems, and provides worked-through examples and exercises from the book in a series of Jupyter notebooks.
posted by Coventry at 11:15 AM on January 4, 2018 [5 favorites]