Profile

Cover photo
Yaroslav Bulatov
Works at Google
Attended Oregon State University
Lives in San Francisco, CA
1,330 followers|305,842 views
AboutPostsPhotosYouTube

Stream

Yaroslav Bulatov

Shared publicly  - 
 
 
I just discovered the computational linguistics olympiad. These problems are a lot of fun. I tried some easy ones and they were easy. So I jumped to a hard one and now I'm stuck. Have fun!

(Don't post any spoilers, I haven't solved it myself yet. Brag about it if you have solved it though!)

http://www.ioling.org
9 comments on original post
5
2
Saravanan Thirumuruganathan's profile photoMarkus Breitenbach's profile photo
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
A map from "Who is afraid of peer review". A fake cancer research paper was sent to 304 author-fee funded journals and 157 accepted. Red on map indicates location of publisher/bank http://scicomm.scimagdev.org/
7
2
Thomas Dietterich's profile photoAdam Mathias's profile photoSteven Van Vaerenbergh's profile photo
 
Too bad they used a globe viewer; this is a case where a flattened map would be much more informative.
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
Deep Learning bibliography of 2014:
http://memkite.com/deep-learning-bibliography/
DeepLearning.University – An Annotated Deep Learning Bibliography. by: Amund Tveit*, Torbjørn Morland and Thomas Brox Røst · memkite.com – Deeplearning.Education. *Email: amund@memkite.com – @atveit. Abstract. DeepLearning.University is an annotated bibliography of recent publications (2014-) ...
9
4
王立君's profile photoC. De Vries's profile photo
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
 
+CVPR 2014  papers list is out !
Learning, recognition, segmentation ... manifold and depth are more and more present.
http://www.pamitc.org/cvpr14/accepted_papers.php
View original post
3
2
alecsandra zbranca's profile photoTroy Lee's profile photoLurk Man's profile photo
 
.
Add a comment...
 
Slides/summaries from last week's Stochastic Gradient Methods workshop: http://yaroslavvb.blogspot.com/2014/03/stochastic-gradient-methods-2014.html
Last week I attended Stochastic Gradient Methods workshop held at UCLA's IPAM . Surprisingly, there's still quite a bit of activity and unsolved questions around what is essentially, minimizing a quadratic function. In 2009 S...
31
20
Hhuiwen Han's profile photoLurk Man's profile photo
Add a comment...
In his circles
599 people
Have him in circles
1,330 people
Larisa Khanarina's profile photo
Evgeni Stavinov's profile photo
Jeff Leffel's profile photo
Ivan Kornienko's profile photo
Sibt ul Hussain's profile photo
mouhamed salem's profile photo
Zoya Barabanova's profile photo
Natasha Martin's profile photo
Nobel Pages's profile photo
 
There's been a recent spate of quantum ML algorithms, but a few pieces are still missing to make them work (even if we had quantum computers) -- Scott Aaronson "Quantum Machine Learning Algorithms: Read the Fine Print" -- http://www.scottaaronson.com/papers/qml.pdf
7
3
David Konerding's profile photoGrzegorz Chrupała's profile photoMohsen Ali's profile photoC. De Vries's profile photo
 
I would so much prefer people prove we can do basic stuff with QC before jumping off to the interesting stuff.
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
 
A secret code

This is the Golay code.  Each row in this picture shows a string of 24 bits.  There are 12 rows.  If you look at any two rows, you'll see they differ in at least 8 places.

Here's how to get the Golay code.  Take a 12 x 12 square of bits with all 0's except for 1's down the diagonal - you can see that at left here.  Take another 12 x 12 square of bits that tells you when two faces of a dodecahedron share an edge: 0 if they do, 1 if they don't.  Stick these squares together and you get the Golay code!

Some guys around here keep asking if the math I talk about is good for anything.  In this case it is! 

The Voyager 1 and 2 spacecraft needed to transmit hundreds of color pictures of Jupiter and Saturn in their 1979, 1980, and 1981 fly-bys.   They had very little bandwidth, so they needed a good error-correcting code.  They used the Golay code! 

The point is that we can use the rows of this picture as code words.  If we take some rows and add them - adding each entry separately, mod 2 - we get more code words.  We get a total of 2^12 = 4096 code words. 

These code words have a cool property: it takes at least 8 errors to turn any code word into any other.   So, we say the Hamming distance between any two code words is at least 8.   In fact, the Golay code is the only code with 12-bit code words where the Hamming distance between any two is at least 8.

There's a whole theory of codes like this, and this is an especially good one.  You can transmit 12 bits of data with 24 bits... but since the Hamming distance between code words is big, someone can understand what you meant even if there are lots of errors!  So, the Golay code is useful for transmitting data in a noisy environment.

But the reason I like the Golay code is that it has a big and important symmetry group.  Its symmetry group is called M24 - one of the amazing things called Mathieu groups.    It has

24 x 23 x 22 x 21 x 20 x 48 = 244,823,040

elements.   It's connected to many other symmetrical things in math: for example, it acts as symmetries of the Leech lattice, the densest way to pack balls in 24 dimensions.

To be more precise, this code here is called the extended binary Golay code.  You can learn more about it here:

https://en.wikipedia.org/wiki/Binary_Golay_code

https://en.wikipedia.org/wiki/Mathieu_group

Puzzle: I said the symmetry group of this code is M24.  But what do I mean, exactly, by a 'symmetry' of this code?

The extended binary Golay code is not only good for outer space.  In 1993, the US government issued standards for high frequency radio systems.  They require using this code for "forwards error correction" in "automatic link establishment"!  See page 51 here:

http://hflink.com/standards/FED_STD_1045A.pdf
47 comments on original post
6
2
Yaroslav Bulatov (Corpacc)'s profile photoYuriy Akopov's profile photo
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
 
I am very proud of the work our team has done on the Eigen open source matrix library this past year. Try out the development (3.3) branch for yourself, you'll likely find that it's the fastest BLAS there is on modern Intel CPUs. Plus it's under a very open license, portable, flexible, and has a very intuitive C++ API.
Eigen 3.2.2 has been released on August 4, 2014. This is a maintenance release with many bug fixes since the release of 3.2.1 five months ago. In particular, this release includes various numerical improvements in JacobiSVD, LDLT, BiCGSTAB, and ColPivHouseholderQR. There are also some limited ...
6 comments on original post
5
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
Stephen Pinker on standardized testing and Ivy League. At Harvard, 5-10% of students are selected based on academic merit. The rest are selected "holistically"

http://www.newrepublic.com/article/119321/harvard-ivy-league-should-judge-students-standardized-tests

"A skilled professional I know had to turn down an important freelance assignment because of a recurring commitment to chauffeur her son to a resumé-building “social action” assignment required by his high school. This involved driving the boy for 45 minutes to a community center, cooling her heels while he sorted used clothing for charity, and driving him back—forgoing income which, judiciously donated, could have fed, clothed, and inoculated an African village. The dubious “lessons” of this forced labor as an overqualified ragpicker are that children are entitled to treat their mothers’ time as worth nothing, that you can make the world a better place by destroying economic value, and that the moral worth of an action should be measured by the conspicuousness of the sacrifice rather than the gain to the beneficiary."
The Ivy League is broken and only standardized tests can fix it.
13
2
Ciprian Chelba's profile photoAndrew Delong's profile photo
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
Summary of last week's Stochastic Gradient Methods workshop: http://yaroslavvb.blogspot.com/2014/03/stochastic-gradient-methods-2014.html
Last week I attended Stochastic Gradient Methods workshop held at UCLA's IPAM . Surprisingly, there's still quite a bit of activity and unsolved questions around what is essentially, minimizing a quadratic function. In 2009 S...
13
9
Jian Zhou's profile photoMeena Mani's profile photo
Add a comment...

Yaroslav Bulatov

Shared publicly  - 
 
Looking forward to seeing how people improve Gradient Descent for neural networks. IPAM's 5-day Stochastic Gradient Methods workshop is in 2 weeks.

https://www.ipam.ucla.edu/programs/sgm2014/
13
2
Dumitru Erhan's profile photoli kev's profile photoYang Chen's profile photo
 
Nice, how it is that I found out via G+ about an stochastic gradient descent event that is happening right in our backyard :). See you in LA in 2 weeks?
Add a comment...
 
Upside of publishing in open review system (for authors at least) -- our paper hasn't been accepted yet, but already picked up by MIT's Technology Review

http://openreview.net/document/0c571b22-f4b6-4d58-87e4-99d7de42a893#0c571b22-f4b6-4d58-87e4-99d7de42a893

+Ian Goodfellow +Julian Ibarz 
Google can identify and transcribe all the views it has of street numbers in France in less than an hour, thanks to a neural network that’s just as good as human operators. Now its engineers reveal how they developed it.
33
6
Troy Lee's profile photoVladimir Bulatov's profile photoBoris Borcic's profile photoSeth Gregory's profile photo
2 comments
 
I'm pretty sure the ArXiv version of my maxout paper gets more citations than the ICML version of it does. It's a double-edged sword: it's good that ArXiv got my work more attention, but it's also bad that most people are evidently finding an older version of the paper without the best final results. The other funny thing is that because I posted the ArXiv paper in the winter, people assume it was an ICLR submission.
Add a comment...
People
In his circles
599 people
Have him in circles
1,330 people
Larisa Khanarina's profile photo
Evgeni Stavinov's profile photo
Jeff Leffel's profile photo
Ivan Kornienko's profile photo
Sibt ul Hussain's profile photo
mouhamed salem's profile photo
Zoya Barabanova's profile photo
Natasha Martin's profile photo
Nobel Pages's profile photo
Work
Employment
  • Google
    Software Engineer, present
    Google Brain
Places
Map of the places this user has livedMap of the places this user has livedMap of the places this user has lived
Currently
San Francisco, CA
Previously
Corvallis, OR - London, England - St.Petersburg, Russia
Story
Tagline
Teaching machines to learn
Introduction
Leningrad>London>Oregon>San Francisco
Education
  • Oregon State University
    Computer Science
Basic Information
Gender
Male
Birthday
July 5, 1980