Data diversity: Preserving variety in data sets should aid machine learning

#1

Post by upen » Sun, Dec 18, 2016, 6:35 PM UTC

Data diversity: Preserving variety in data sets should aid machine learning
news.mit.edu

Re: Data diversity: Preserving variety in data sets should aid machine learning

#2

Post by q_revert » Sun, Dec 18, 2016, 7:45 PM UTC

I think this is the paper, which oddly isn't linked in the article:

https://arxiv.org/abs/1509.01618

Re: Data diversity: Preserving variety in data sets should aid machine learning

#3

Post by opaqe » Sun, Dec 18, 2016, 7:48 PM UTC

Is there a more detailed paper describing the algorithm? The description is very vague in the article. When they pick the two points, is there an evaluation on how much "diversity" increases w/r/t each of the three possible operations, and that's how they choose?

edit: thanks @q_revert for linking the paper

Re: Data diversity: Preserving variety in data sets should aid machine learning

#4

Post by sidrajaram » Sun, Dec 18, 2016, 9:28 PM UTC

q_revert wrote: post #2

I think this is the paper, which oddly isn't linked in the article: https://arxiv.org/abs/1509.01618

That seems to be a precursor to the work mentioned in the article. This is the one that was presented at NIPS this year: https://papers.nips.cc/paper/6182-fast-mixing-markov-chains-...

Re: Data diversity: Preserving variety in data sets should aid machine learning

#5

Post by rokosbasilisk » Sun, Dec 18, 2016, 9:45 PM UTC

I believe they are using mcmcs at the core. markov chain multi carlos. this might be useful if you are wondering what it is http://mlwhiz.com/blog/2015/08/19/MCMC_Algorithms_Beta_Distr...

Re: Data diversity: Preserving variety in data sets should aid machine learning

#6

Post by ultrafilter » Sun, Dec 18, 2016, 10:17 PM UTC

q_revert wrote: post #2

I think this is the paper, which oddly isn't linked in the article: https://arxiv.org/abs/1509.01618

The article has a link in a sidebar:

https://arxiv.org/abs/1608.01008v2