Would like to actually see these CNN's first hand but if I'm not missing something on a first quick pass, the amazing results of 2013-2014 are continuing. Major gating barriers that have stymied AGI's are disappearing. This result is cool on its own. But, what really excites is how this provides a key foundational building block for even cooler ideas. When you have sentence structure, it naturally lends itself to many unsupervised NLP and learning tasks. Just wow!
Word2Vec is getting referenced a lot these days, but that doesn't mean much about the referencing paper. This approach is pretty unique in trying to apply CNNs (which are very successful in image recognition) to NLP. But it's not yet clear to me whether these are just tricks or are they really onto some key insight.
The back-and-forth on the ImageNet record between Google/Facebook/Baidu suggests that, unless they're exquisitely coordinating research release, at least some of what they're releasing is indeed close to their state of the art.
obviously writing it up does mean the specified results will be a bit behind what they can actually do in their labs, but that's true of everywhere
Not sure why that follows. (Can you outline the reasoning a bit more?)
As an example model, imagine all three are years ahead of published material, and each has policies that encourage publishing only the minimum necessary to "take the crown" (because any more would risk diluting proprietary advantages).
That'd result in the observed back-and-forth, and – given lead-times on paper-writing, internal review, and marquee conferences – it wouldn't necessarily force an acceleration of the "published state-of-the-art" up to the level of all of their "internal states-of-the-art". (Perhaps, for a group in trailing/catch-up position, their published results will be very close to their best. But the leader could be arbitrarily further ahead.)
By analogy to English auction bidding: outsiders only learn the second-highest reserve price just before the end, and never learn the winner's true reserve price. But here there's no end, and there's always potential competitive reasons for the top-N pack to hide some of their leading-edge practices.
This doesn't require explicit coordination… but there could be explicit coordination, too! All the programs are staffed by former students/colleagues/coworkers of each other.
> That'd result in the observed back-and-forth, and – given lead-times on paper-writing, internal review, and marquee conferences – it wouldn't necessarily force an acceleration of the "published state-of-the-art" up to the level of all of their "internal states-of-the-art".
If they aren't coordinating very carefully, a single defector would result in the entire slack being used up by a single announcement; and the bigger the slack used up, the more a PR win it is...
OK, but in this sort of race, publishing your very latest internal techniques lets everyone else instantly catch up.
If these techniques are commercially valuable and in current use – and I believe they are! – then no leading company, or even member of the leading pack, would want to do a current-best reveal. They all have competitive reasons to do only carefully vetted, incremental reveals of somewhat-older work.
You can publish papers without giving away all the details necessary to make it work. Sad but true.
For example, submitted to HN sometime ago was the blog of a dude trying to make Deepmind's 'neural turing machine' work; he was having a hella time because the published paper seems to be unclear or skip over a number of crucial points. Or more historically, the German chemical giants made an art of filing patents on all their key techniques, to gain IP protection, but leaving out enough crucial details that when the USA gleefully seized their IP rights during WWI, the American companies discovered they couldn't make the processes work.
Indeed, and if that's also happening with the deep-learning papers here, then it's another mechanism in support of my main point: what's published is often, by motivated choice, behind what's being done internally. (And not just, "time it takes to write up" behind.)
Given that intent, there's no way to deduce from the pattern of new claimed results whether what's being revealed is a merely a few months, or many years, behind.
How do you guys understand this stuff!?
The topic absolutely fascinated me, but after reading the paper... I just feel dumb. Are there any good resources I could use to better understand machine learning papers such as this? I mean I can't even comprehend the implications this paper could have? Is anyone opening to doing some mentoring?
Gbachik@gmail.com
Machine learning can be quite jargon-heavy, and in my experience the core ideas can often be hidden amongst a bunch of unnecessary maths.
I got frustrated by a paper yesterday which contained function definitions, summations-of-summations, products of sequences, convolutions, set theory, switching back-and-forth between unary-functions/vectors and binary-functions/matrices, converting back-and-forth between {0, 1}, {-1, 1} and {true, false}, weighting elements of a set by 0/1 instead of taking a sub-set, linear programming, etc.
What was their result? To speed up pair-wise comparisons of structured data, only do N% of the comparisons and it will only take N% of the time. To decide which comparisons to discard, see what works well on a small sample of inputs.
The progress here is getting spooky, including the success of such networks for captioning arbitrary images and translating between natural languages. I read that one researcher predicts live narration of video within 5 years.
A cold-splash-in-the-face intro for laypeople can be found in the TEDx talk of Jeremy Howard (founder of Kaggle):
There's some pretty good stuff there. I really liked Crypto-Nets: Neural Networks over Encrypted Data[1]:
The problem we address is the following: how can a user employ a predictive model that is held by a third party, without compromising private information. For example, a hospital may wish to use a cloud service to predict the readmission risk of a patient. However, due to regulations, the patient's medical files cannot be revealed. The goal is to make an inference using the model, without jeopardizing the accuracy of the prediction or the privacy of the data.
Why don't we reallocate some of this effort to teaching humans how to use language more effectively? Or is it better to just delegate the responsibility for truth to a seemingly "pure" logic that we like to think exists outside of our human condition?
If you see convolution as basically truncated recurrence this approach ties in very strongly to recent approaches to machine translation using recurrent nets. I guess depth should allow you to find longterm dependencies, but the fact that CNN were designed for images which have strong local structure and much weaker long term structure makes me think RNNs are better for language, where we see a lot of important long term dependencies. As an example: "The man with the long brown hair entered the saloon" - I would tie saloon and man as the key pieces of that sentence, but that dependency is pretty long and somewhat different than natural images where you don't really expect the corners of images to have any strong relationship in general.
^agreed - no ANN can be treated as a 'jack of all trades' because of the success found in one domain (it seems more networks are designed to boost performance on one type of data)
24 comments
[ 4.6 ms ] story [ 94.8 ms ] threadMy hunch (as a total outsider) is that anything Google publishes is about 2 years behind their current best practices.
obviously writing it up does mean the specified results will be a bit behind what they can actually do in their labs, but that's true of everywhere
As an example model, imagine all three are years ahead of published material, and each has policies that encourage publishing only the minimum necessary to "take the crown" (because any more would risk diluting proprietary advantages).
That'd result in the observed back-and-forth, and – given lead-times on paper-writing, internal review, and marquee conferences – it wouldn't necessarily force an acceleration of the "published state-of-the-art" up to the level of all of their "internal states-of-the-art". (Perhaps, for a group in trailing/catch-up position, their published results will be very close to their best. But the leader could be arbitrarily further ahead.)
By analogy to English auction bidding: outsiders only learn the second-highest reserve price just before the end, and never learn the winner's true reserve price. But here there's no end, and there's always potential competitive reasons for the top-N pack to hide some of their leading-edge practices.
This doesn't require explicit coordination… but there could be explicit coordination, too! All the programs are staffed by former students/colleagues/coworkers of each other.
If they aren't coordinating very carefully, a single defector would result in the entire slack being used up by a single announcement; and the bigger the slack used up, the more a PR win it is...
If these techniques are commercially valuable and in current use – and I believe they are! – then no leading company, or even member of the leading pack, would want to do a current-best reveal. They all have competitive reasons to do only carefully vetted, incremental reveals of somewhat-older work.
For example, submitted to HN sometime ago was the blog of a dude trying to make Deepmind's 'neural turing machine' work; he was having a hella time because the published paper seems to be unclear or skip over a number of crucial points. Or more historically, the German chemical giants made an art of filing patents on all their key techniques, to gain IP protection, but leaving out enough crucial details that when the USA gleefully seized their IP rights during WWI, the American companies discovered they couldn't make the processes work.
Given that intent, there's no way to deduce from the pattern of new claimed results whether what's being revealed is a merely a few months, or many years, behind.
http://deeplearning.net/tutorial/
And this book (work in progress):
http://www.iro.umontreal.ca/~bengioy/dlbook/
[1]: Wiki with code, exercises and explanation
[2]: Video lecture one with a recap on back-propagation
[3]: Video lecture two on Sparse Auto Encoders
[4]: Handouts
[1]: http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial
[2]: http://www.stanford.edu/class/cs294a/video1.html
[3]: http://www.stanford.edu/class/cs294a/video2.html
[4]: http://www.stanford.edu/class/cs294a/handouts.html
I got frustrated by a paper yesterday which contained function definitions, summations-of-summations, products of sequences, convolutions, set theory, switching back-and-forth between unary-functions/vectors and binary-functions/matrices, converting back-and-forth between {0, 1}, {-1, 1} and {true, false}, weighting elements of a set by 0/1 instead of taking a sub-set, linear programming, etc.
What was their result? To speed up pair-wise comparisons of structured data, only do N% of the comparisons and it will only take N% of the time. To decide which comparisons to discard, see what works well on a small sample of inputs.
A cold-splash-in-the-face intro for laypeople can be found in the TEDx talk of Jeremy Howard (founder of Kaggle):
The wonderful and terrifying implications of computers that can learn – https://www.youtube.com/watch?v=xx310zM3tLs
For cutting-edge research, it seems the "NIPS" conference each December is where many of the new results appear:
http://nips.cc/
There's some pretty good stuff there. I really liked Crypto-Nets: Neural Networks over Encrypted Data[1]:
The problem we address is the following: how can a user employ a predictive model that is held by a third party, without compromising private information. For example, a hospital may wish to use a cloud service to predict the readmission risk of a patient. However, due to regulations, the patient's medical files cannot be revealed. The goal is to make an inference using the model, without jeopardizing the accuracy of the prediction or the privacy of the data.
[1] http://arxiv.org/abs/1412.6181