ML Newsnew | past | best | rss | submit about
Deep Gate Recurrent Neural Network (arxiv.org)
50 points by vonnik on May 20, 2016 | hide | past | pdf | 8 comments on HN

In plain words: A new memory cell for learning long patterns in data uses one gate to decide what to keep, where the usual cells use several gates to control the flow. It needs fewer parameters and less computation, and learned long-range patterns faster than those cells.

Abstract

This paper introduces two recurrent neural network structures called Simple Gated Unit (SGU) and Deep Simple Gated Unit (DSGU), which are general structures for learning long term dependencies. Compared to traditional Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), both structures require fewer parameters and less computation time in sequence classification tasks. Unlike GRU and LSTM, which require more than one gates to control information flow in the network, SGU and DSGU only use one multiplicative gate to control the flow of information. We show that this difference can accelerate the learning speed in tasks that require long dependency information. We also show that DSGU is more numerically stable than SGU. In addition, we also propose a standard way of representing inner structure of RNN called RNN Conventional Graph (RCG), which helps analyzing the relationship between input units and hidden units of RNN.

Yuan Gao, Dorota Glowacka
arXiv:1604.02910 · cs.NE · submitted Apr 11, 2016 · updated May 13, 2016
abstract · pdf · This paper has been withdrawn by the author due to lacking of enough experiments

add comment on HN

jean- on May 21, 2016 [–]
The comment field on arXiv reads:

> This paper has been withdrawn by the author due to lacking of enough experiments

reply
argonaut on May 21, 2016 [–]
PSA: In case you didn't know, anyone can publish anything to arxiv.
reply
colah3 on May 21, 2016 | parent [–]
I missed the actual paper before it got taken down. But as someone in the field, I'd like to say that these are real researchers at a respected group.

The Helsinki group has done a lot of lovely work over the last few years -- ladder nets come to mind. This paper is by Yuan Gao, and his advisor Dorota Glowacka. Dr. Glowacka has a pretty substantial publication record, mostly focused on reinforcement learning:

https://scholar.google.com/citations?user=sDZkDHQAAAAJ&hl=en

Yuan Gao is a second year grad student, and this is one of his first publications. I think it shows a lot of integrity to realize a weakness in your work and remove it, pending refinement.

reply
Djabbz on May 21, 2016 | parent [–]
That's exactly why double blind reviewing is necessary. #bias
reply
colah3 on May 21, 2016 | parent [–]
I hope my previous comment didn't come across as me suggesting that the paper shouldn't be scrutinized because it comes from an established research group. Of course, it absolutely should be and double blind review is a helpful mechanism for achieving this.

One possible reading of the parent comment was an attack on the researchers. Of course, the parent probably didn't mean it that way but it might still feel hurtful to the authors and I wanted to clarify it. :)

reply
argonaut on May 21, 2016 | parent [–]
It was an attack on HN'ers blindly upvoting things from arxiv, because most HN readers do not properly scrutinize postings from arxiv or take it with a grain of salt.

I was well aware these were actual researchers. Your comment probably didn't help very much here because most HN'ers, again, will just see that the authors are PhD's or from a decent group, and assume it's a legit paper.

reply
zk00006 on May 21, 2016 [–]
where is the link to pdf?
reply
Houshalter on May 21, 2016 | parent [–]
http://arxiv.org/pdf/1604.02910v2
reply
About | RSS | RSS (all) | HN arXiv