Profile

Cover photo
Cosmin Negruseri
1,267 followers|516,973 views
AboutPostsReviews

Stream

Cosmin Negruseri

Shared publicly  - 
 
Chefchauen
6
Add a comment...

Cosmin Negruseri

Shared publicly  - 
 
 
Google regains the Imagenet (ILSVRC 2012) classification mantle, but just barely: 4.82% top-5 classification error vs 4.94% for Microsoft. Best human accuracy on this task is about 5% error. 

What's really cool is the batch normalization idea in the paper. It is not only a huge speed improvement, but also a notable quality improvement. This should become the new standard way of training deep nets. Worth noting is that the Inception net we trained is much cheaper than the recent Microsoft or Baidu counterparts. 

Congratulations +Sergey Ioffe  and +Christian Szegedy on this awesome work. 
Abstract: Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, ...
View original post
1
Add a comment...

Cosmin Negruseri

Shared publicly  - 
9
1
Andrei Savu's profile photo
Add a comment...

Cosmin Negruseri

Shared publicly  - 
7
Add a comment...

Cosmin Negruseri

Shared publicly  - 
 
Here's a neat problem I've heard from Christian Szegedy:

N cars are moving in the same direction with different speeds on an infinite straight road. The cars can't pass eachother so clusters form. What's the expected number of clusters?
4
Octavian Costache's profile photoCosmin Negruseri's profile photoRadu Grigore's profile photo
12 comments
 
I said log n without doing any analysis because I thought it's meant to be equivalent to counting assignments in "m=-inf; for i in 1 to n; if a[i] > m then m=a[i]" which in turn is equivalent to counting cycles in permutations. I knew this problem is analysed to death in TAoCP and the answer is H(n), and I thought log(n) is good enough as a hint :p. (Then I saw the problem was solved on infoarena before I post my comment, and Radu Berinde already mentioned explicitly the connection to random permutations.)

I looked it up now: sections 1.2.10 and 1.3.3 (towards end). Knuth also gives variance H(n,1)-H(n,2). You could ask on infoarena what's the variance ;)

Which reminds me: there are some interesting observations in the discussion there (the variant in which a lonely care doesn't make a cluster, how to quickly estimate log n without going thru H(n), length of longest cluster) so a link might be good http://www.infoarena.ro/blog/cars
Add a comment...

Cosmin Negruseri

Shared publicly  - 
 
South Kaibab trail - Phantom Ranch - Bright Angel trail.
28km, 1700m, 10 hours
7
Add a comment...
Have him in circles
1,267 people
Castiga .net's profile photo
Zhanibek Datbayev's profile photo
Tudor Damian (Tudy)'s profile photo
Bob Garlick's profile photo
Uniristii CNU Focsani's profile photo
crysty draghici's profile photo
Hamid Amallah's profile photo
Adi Bura's profile photo
Bogdan Strugari's profile photo

Cosmin Negruseri

Shared publicly  - 
5
Add a comment...

Cosmin Negruseri

Shared publicly  - 
3
Add a comment...

Cosmin Negruseri

Shared publicly  - 
8
Add a comment...

Cosmin Negruseri

Shared publicly  - 
 
Kauai
3
Add a comment...

Cosmin Negruseri

Shared publicly  - 
 
My friend Christian is making a case that Rectified Linear Units are responsible for the recent deep learning revolution in his talk about the winning entry in the ImageNet 2014 winning entry.
ILSVRC2014: GoogLeNet
2
Add a comment...
People
Have him in circles
1,267 people
Castiga .net's profile photo
Zhanibek Datbayev's profile photo
Tudor Damian (Tudy)'s profile photo
Bob Garlick's profile photo
Uniristii CNU Focsani's profile photo
crysty draghici's profile photo
Hamid Amallah's profile photo
Adi Bura's profile photo
Bogdan Strugari's profile photo
Links
Other profiles
Contributor to
Basic Information
Gender
Male
9 reviews