5 ms·

An Adversarial Review of “Adversarial Generation of Natural Language”

119 points by sebg 9y ago

slashcom 9y ago
It's worth noting this post caused a very intense twitter debate among the NLP and DL communities, especially after Yann Lecun replied to Yoav's comments. https://www.facebook.com/yann.lecun/posts/10154498539442143 https://www.facebook.com/yann.lecun/posts/10154498539442143
- sillysaurus3 9y ago
  What are some twitter links? I'd be interested to see the kind of nuanced conversations that twitter can spawn. Also looking to follow some NLP people.
  - Mathnerd314 9y ago
    Twitter has a search box: https://twitter.com/search?f=tweets&vertical=default&q=An%20Adversarial%20Review%20of%20%E2%80%9CAdversarial%20Generation%20of%20Natural%20Language%E2%80%9D https://twitter.com/search?f=tweets&vertical=default&q=An%20... This newsletter looks interesting (mentioned on Twitter): http://digest.deeplearningweekly.com/issues/deep-learning-weekly-issue-44-ai-understanding-google-mobilenet-machine-intelligence-summits-sobolev-trainings-for-nns-flag-planting-and-overselling-results-61446 http://digest.deeplearningweekly.com/issues/deep-learning-we... This seems like the only real discussion from the search: https://twitter.com/kchonyc/status/873305485428088833 https://twitter.com/kchonyc/status/873305485428088833 I guess search is broken, the author's twitter feed has a bunch more discussion (https://twitter.com/yoavgo/with_replies https://twitter.com/yoavgo/with_replies): https://twitter.com/yoavgo/status/872831207163265024 https://twitter.com/yoavgo/status/872831207163265024 https://twitter.com/yoavgo/status/872968874521702400 https://twitter.com/yoavgo/status/872968874521702400 https://twitter.com/Smerity/status/872971766959718400 https://twitter.com/Smerity/status/872971766959718400 https://twitter.com/yoavgo/status/873489277157507072 https://twitter.com/yoavgo/status/873489277157507072 https://twitter.com/jacobandreas/status/873109327644573696 https://twitter.com/jacobandreas/status/873109327644573696 https://twitter.com/yoavgo/status/873175136056336386 https://twitter.com/yoavgo/status/873175136056336386 https://twitter.com/hugo_larochelle/status/873369688675430400 https://twitter.com/hugo_larochelle/status/87336968867543040... https://twitter.com/kchonyc/status/873306255204507648 https://twitter.com/kchonyc/status/873306255204507648 https://twitter.com/yoavgo/status/873786844315607040 https://twitter.com/yoavgo/status/873786844315607040 https://twitter.com/haldaume3/status/873565061754781697 https://twitter.com/haldaume3/status/873565061754781697
    - sp332 9y ago
      None of those tweets have the phrase you searched for, so why do you conclude that search is broken?
      - Mathnerd314 9y ago
        Several of the tweets link to the post, and the title shows up in Twitter's preview box (although not in the tweet proper). If Twitter search "worked" it should have found those, via some magic string munging or whatever. My definition of "works" is https://en.wikipedia.org/wiki/DWIM https://en.wikipedia.org/wiki/DWIM, others are free to disagree.
fizixer 9y ago
This is an instance of the general issue of conflict across, what I like to call, the salesperson-slacker spectrum. Most researchers/academics lie somewhere on this spectrum. (Well I guess most human beings involved in any activity probably). On the one end are salespeople who love to make a mountain out of a molehill they just discovered. On the other, slackers are like the perfectionists who never get anything done because they never resolved their analysis-paralysis. There are very few who are exactly in the middle of the spectrum. The middle is a point of unstable equilibrium. You have to work very hard to stay there and can easily fall off to one side or the other.
- tibbetts 9y ago
  Your model fails to admit the possibility of researchers who do better or worse work.
  - fizixer 9y ago
    Model fixed ('every' changed to 'most').
- obstinate 9y ago
  I wouldn't say perfectionism and sales-orientedness are on the same dimension. However, I have noticed that being given the hard-sell is more often than not associated with the product being sub par. A product or project that is doing very good work often doesn't have to sell quite as hard, because the work speaks for itself. (My experience on this issue mostly deals with internal teams. I don't know how much this applies to external teams.) This is especially true when two products have teams of about equal size, experience, and expertise. Time spent selling is time not spent making a better product. All other things being equal, every hour spent selling is one less hour spent working on making your product better.
  - Swizec 9y ago
    And with exactly zero selling you have the best product you can make but one that nobody knows exists.
a3_nm 9y ago
> This post is also an ideological action w.r.t arxiv publishing Anyone else thought that this was very weird? The author appears to be complaining about the fact that reputable people/labs can post a PDF on arXiv and be taken seriously. How is this avoidable? Without arXiv, they could just post the PDF on their website or anywhere else. The "risk" associated to publishing crap on arXiv is the same as always: have people notice it's crap and get a bad reputation. I'm not sure what ideology has to do with it.
- denzil_correa 9y ago
  > Anyone else thought that this was very weird? The author appears to be complaining about the fact that reputable people/labs can post a PDF on arXiv and be taken seriously. How is this avoidable? Without arXiv, they could just post the PDF on their website or anywhere else. The issue is "flag planting" : arXiv is one way to do it; the same arguments hold true for other sources but arXiv is used as a canonical example for "flag planting".
- naturalgradient 9y ago
  No, this is not what this is about at all. We have reached a point where reviewers of reputable conferences will ask why a paper is not referencing unreviewed work that has been uploaded on arxiv a day before the conference deadline. This is not hyperbole, as anyone who is submitting to ICLR or NIPS can confirm. Work of certain labs is taken as established authority as soon as it hits arxiv.
- Al-Khwarizmi 9y ago
  Indeed it's possible in theory to do flag planting without arxiv, for example by posting a timestamped technical report to an institutional repository, but it has never been a common thing. Probably because you don't get much of an audience, so people won't cite you unless you bug them in reviews. And if someone publishes the idea some months later in a paper, they will probably get credit as there's the reasonable assumption they just hadn't read your report. On the other hand, arxiv has a huge audience, lately maybe even more than DL/NLP conferences, making the flag planting really effective, especially if you are from a prestigious group. So there is a real problem now with large, prestigious groups posting half-assed preliminary results in arxiv, and deterring more modest groups from working on the problem or bogging them down because they now need to compare themselves to the well-known arxiv approach, which often has serious reproducibility issues because it has not been peer-reviewed. I'm not an arxiv hater, in fact I check arxiv every day and for the last year or so I have posted most of my papers in it. But the problem is real and something must be done. Not about arxiv which is just the messenger (and a good one), but about the flag-planting culture using it that has emerged in the field.
paganel 9y ago
I'm not going to comment about that paper being published on arxiv and I don't generally care about the NLP vs DL debate, just wanted to say that those generated examples did look indeed all rubbish to me. As do most of the Google Translate pieces, even though I get the feeling that automatic translation of texts is now seen almost as a solved problem (it's not): all that Google Translate does is change some text from some original language to a second one, which is not a real language, it's just a language that's sometimes very close (grammatically and lexical) to one which the agent/user knows. The idea is that we should try to look harder and have fairer judgements about the actual results and not get stuck on the methodologies.
denzil_correa 9y ago
> Communities will naturally recognize contributions and give credit when credit is due. It's always happen that way. "Let the market decide!"
- mostafab 9y ago
  The market needs signals, and the point of Goldberg was that those signals are missing with arxiv, and that just looking at the lab reputation was a poor signal. I am working on it, contact me for details.
wodenokoto 9y ago
If you want to read more discussion on this topic (and this article), see the article "A computational linguistic farce in 3 acts" and its HN discussion: https://news.ycombinator.com/item?id=14532306 https://news.ycombinator.com/item?id=14532306
thearn4 9y ago
For another point of comparison, has there been a "us vs. them" dynamic in the computer vision community when it comes to deep learning? After all, it seems like deep convolutional networks did sort of railroad there way through a ton of topics in that field. I used to do image analysis research, but moved on before much of this came into play so I lack any sort of inside scoop.

119 points by sebg 9y ago

It's worth noting this post caused a very intense twitter debate among the NLP and DL communities, especially after Yann Lecun replied to Yoav's comments. https://www.facebook.com/yann.lecun/posts/10154498539442143 https://www.facebook.com/yann.lecun/posts/10154498539442143

sillysaurus3 9y ago
What are some twitter links? I'd be interested to see the kind of nuanced conversations that twitter can spawn. Also looking to follow some NLP people.
- Mathnerd314 9y ago
  Twitter has a search box: https://twitter.com/search?f=tweets&vertical=default&q=An%20Adversarial%20Review%20of%20%E2%80%9CAdversarial%20Generation%20of%20Natural%20Language%E2%80%9D https://twitter.com/search?f=tweets&vertical=default&q=An%20... This newsletter looks interesting (mentioned on Twitter): http://digest.deeplearningweekly.com/issues/deep-learning-weekly-issue-44-ai-understanding-google-mobilenet-machine-intelligence-summits-sobolev-trainings-for-nns-flag-planting-and-overselling-results-61446 http://digest.deeplearningweekly.com/issues/deep-learning-we... This seems like the only real discussion from the search: https://twitter.com/kchonyc/status/873305485428088833 https://twitter.com/kchonyc/status/873305485428088833 I guess search is broken, the author's twitter feed has a bunch more discussion (https://twitter.com/yoavgo/with_replies https://twitter.com/yoavgo/with_replies): https://twitter.com/yoavgo/status/872831207163265024 https://twitter.com/yoavgo/status/872831207163265024 https://twitter.com/yoavgo/status/872968874521702400 https://twitter.com/yoavgo/status/872968874521702400 https://twitter.com/Smerity/status/872971766959718400 https://twitter.com/Smerity/status/872971766959718400 https://twitter.com/yoavgo/status/873489277157507072 https://twitter.com/yoavgo/status/873489277157507072 https://twitter.com/jacobandreas/status/873109327644573696 https://twitter.com/jacobandreas/status/873109327644573696 https://twitter.com/yoavgo/status/873175136056336386 https://twitter.com/yoavgo/status/873175136056336386 https://twitter.com/hugo_larochelle/status/873369688675430400 https://twitter.com/hugo_larochelle/status/87336968867543040... https://twitter.com/kchonyc/status/873306255204507648 https://twitter.com/kchonyc/status/873306255204507648 https://twitter.com/yoavgo/status/873786844315607040 https://twitter.com/yoavgo/status/873786844315607040 https://twitter.com/haldaume3/status/873565061754781697 https://twitter.com/haldaume3/status/873565061754781697
  - sp332 9y ago
    None of those tweets have the phrase you searched for, so why do you conclude that search is broken?
    - Mathnerd314 9y ago
      Several of the tweets link to the post, and the title shows up in Twitter's preview box (although not in the tweet proper). If Twitter search "worked" it should have found those, via some magic string munging or whatever. My definition of "works" is https://en.wikipedia.org/wiki/DWIM https://en.wikipedia.org/wiki/DWIM, others are free to disagree.

fizixer 9y ago

This is an instance of the general issue of conflict across, what I like to call, the salesperson-slacker spectrum. Most researchers/academics lie somewhere on this spectrum. (Well I guess most human beings involved in any activity probably). On the one end are salespeople who love to make a mountain out of a molehill they just discovered. On the other, slackers are like the perfectionists who never get anything done because they never resolved their analysis-paralysis. There are very few who are exactly in the middle of the spectrum. The middle is a point of unstable equilibrium. You have to work very hard to stay there and can easily fall off to one side or the other.

tibbetts 9y ago
Your model fails to admit the possibility of researchers who do better or worse work.
- fizixer 9y ago
  Model fixed ('every' changed to 'most').
obstinate 9y ago
I wouldn't say perfectionism and sales-orientedness are on the same dimension. However, I have noticed that being given the hard-sell is more often than not associated with the product being sub par. A product or project that is doing very good work often doesn't have to sell quite as hard, because the work speaks for itself. (My experience on this issue mostly deals with internal teams. I don't know how much this applies to external teams.) This is especially true when two products have teams of about equal size, experience, and expertise. Time spent selling is time not spent making a better product. All other things being equal, every hour spent selling is one less hour spent working on making your product better.
- Swizec 9y ago
  And with exactly zero selling you have the best product you can make but one that nobody knows exists.

a3_nm 9y ago

> This post is also an ideological action w.r.t arxiv publishing Anyone else thought that this was very weird? The author appears to be complaining about the fact that reputable people/labs can post a PDF on arXiv and be taken seriously. How is this avoidable? Without arXiv, they could just post the PDF on their website or anywhere else. The "risk" associated to publishing crap on arXiv is the same as always: have people notice it's crap and get a bad reputation. I'm not sure what ideology has to do with it.

denzil_correa 9y ago
> Anyone else thought that this was very weird? The author appears to be complaining about the fact that reputable people/labs can post a PDF on arXiv and be taken seriously. How is this avoidable? Without arXiv, they could just post the PDF on their website or anywhere else. The issue is "flag planting" : arXiv is one way to do it; the same arguments hold true for other sources but arXiv is used as a canonical example for "flag planting".
naturalgradient 9y ago
No, this is not what this is about at all. We have reached a point where reviewers of reputable conferences will ask why a paper is not referencing unreviewed work that has been uploaded on arxiv a day before the conference deadline. This is not hyperbole, as anyone who is submitting to ICLR or NIPS can confirm. Work of certain labs is taken as established authority as soon as it hits arxiv.
Al-Khwarizmi 9y ago
Indeed it's possible in theory to do flag planting without arxiv, for example by posting a timestamped technical report to an institutional repository, but it has never been a common thing. Probably because you don't get much of an audience, so people won't cite you unless you bug them in reviews. And if someone publishes the idea some months later in a paper, they will probably get credit as there's the reasonable assumption they just hadn't read your report. On the other hand, arxiv has a huge audience, lately maybe even more than DL/NLP conferences, making the flag planting really effective, especially if you are from a prestigious group. So there is a real problem now with large, prestigious groups posting half-assed preliminary results in arxiv, and deterring more modest groups from working on the problem or bogging them down because they now need to compare themselves to the well-known arxiv approach, which often has serious reproducibility issues because it has not been peer-reviewed. I'm not an arxiv hater, in fact I check arxiv every day and for the last year or so I have posted most of my papers in it. But the problem is real and something must be done. Not about arxiv which is just the messenger (and a good one), but about the flag-planting culture using it that has emerged in the field.

paganel 9y ago

I'm not going to comment about that paper being published on arxiv and I don't generally care about the NLP vs DL debate, just wanted to say that those generated examples did look indeed all rubbish to me. As do most of the Google Translate pieces, even though I get the feeling that automatic translation of texts is now seen almost as a solved problem (it's not): all that Google Translate does is change some text from some original language to a second one, which is not a real language, it's just a language that's sometimes very close (grammatically and lexical) to one which the agent/user knows. The idea is that we should try to look harder and have fairer judgements about the actual results and not get stuck on the methodologies.

denzil_correa 9y ago

> Communities will naturally recognize contributions and give credit when credit is due. It's always happen that way. "Let the market decide!"

mostafab 9y ago
The market needs signals, and the point of Goldberg was that those signals are missing with arxiv, and that just looking at the lab reputation was a poor signal. I am working on it, contact me for details.

wodenokoto 9y ago

If you want to read more discussion on this topic (and this article), see the article "A computational linguistic farce in 3 acts" and its HN discussion: https://news.ycombinator.com/item?id=14532306 https://news.ycombinator.com/item?id=14532306

thearn4 9y ago

For another point of comparison, has there been a "us vs. them" dynamic in the computer vision community when it comes to deep learning? After all, it seems like deep convolutional networks did sort of railroad there way through a ton of topics in that field. I used to do image analysis research, but moved on before much of this came into play so I lack any sort of inside scoop.