Google applying to patent deep neural network (LSTM) for machine translation
freepatentsonline.com
freepatentsonline.com
(didn't dig too far into this but..) why y'all need to patent this then?
Patents can be beneficial to facilitate constructive competition, but think humanity is best served by neural nets becoming the new electricity rather than the new Apple-esque walled-garden...
so someone else doesn't and trolls with it?
In this case there is a provisional patent from 2014, and this application follows from that provisional.
It's a permanent record, but it's not published. This is not a great plan.
In the US you get a year after public disclosure to file even a provisional. But not so if you want non-US patents. So you are closing some doors but keeping others open with public disclosure. Also, you are starting a 1 year clock.
A provisional is private; it does not count as public disclosure. A provisional is nothing more than a priority date, assuming what you have disclosed in the provisional itself is sufficient, and novel. You can even refile the same provisional every year as long as you believe the subject mater is still novel, but you get a new filing/priority date each time.
As for preventing a 3rd party from patenting the subject matter, either one is sufficient. However, if you publically disclose, only then you also get protection from a 3rd party who builds on your work. So in that case public disclosure is better than the 'secret' provisional.
Better yet, just timestamp the document in the blockchain.
So best case scenario, Google doesn't want to be caught with its pants down regarding patents. Worst case, it wants to "own" deep learning, so that nobody can really compete with them. Although I think that would be a little in conflict with their strategy to open source tensorflow.
To really figure out on which side Google is now playing we'll have to see how they respond to future patent reforms, and whether they join Microsoft and IBM to once again kill those reforms, or support the reforms to abolish software patents or drastically reduce their damage.
Only a little. Releasing all the models and frameworks helps advance the field, helps with finding people to recruit, helps with integrating them into teams, and so on. This is why so many giants find it in their own self-interest to contribute to FLOSS these days.
Competition-wise, as is often said, Google has all the data. If for every deep learning advance they make $1 and the competitors make $0.95, they win. Patents here are quite helpful: you may make a neat translation app using some new tricks, and then discover when you go to commercialize it that oops, Google's patented 'using neural nets for translation'. Then you either quit, get sued, get bought, or give them most of your profits.
If you patent the core idea, the other patents become a lot less useful. (Not that I think Google is thinking this way. It's just a PR problem to them. When no one is looking, I bet they do whatever they can to get as much money/power as they can.)
FTFY
http://arstechnica.com/business/2007/03/analysis-microsofts-...
One quick change in business strategy could turn Google into the world's largest patent troll.
Is there any big tech company for which this statement does not apply? Or do you hold Google to a higher standard?
I, on the other hand, consider Google holding a patent not significantly different than Oracle holding a patent.
I had a discussion about this with an IBM representative 15 years ago at a symposium at Heinrich-Boell-Stiftung in Berlin: IBM's point was that one can always join their patent-pool (of defensive patents) which means you give them a free license to use your patents and you are free to use their "defensive" patents. This is completely broken: You never know which patents become relevant and any new player already lost because one just cannot keep up with a company that can extort a free license from you and then dump a few (hundred) million into development based on that.
This is a patent on a very particular form of translation model that handles rare words, i.e. this paper that all the authors are on: http://arxiv.org/abs/1410.8206
If you open source something, there should be a reasonable expectation that it is contributed to the commons. Otherwise a lot of people will build off your stuff, which you can then turn around, file a patent, and claim infringement.
Not a lawyer, but seems like a problem.
I think that even though the application is for a specific architecture, that this is still worth knowing about, since LSTM is a such well known technique for dealing with sequential data like sentences, and since so far virtually all progress in AI and ML has been driven by academia and has remained open.
I also think it's important to note that pretty much any interesting machine translation task will contain rare words, so even though they've framed it in terms of a specific task, it's one that's actually extremely broad. Machine translation is not so hard when you have tons of data and when you've seen every word many times in combination with its translations. The only really interesting case is when we have to use background knowledge and context to infer meaning. Since natural languages are notoriously ambiguous, this happens all the time. So this app may be broader than it first appears.
Also, the concept of LSTM wouldn't be patentable directly even in the current liberal software patent interpretation, you might patent particular applications of LSTM (e.g. this patent) but not any and all arbitrary applications of LSTM.