These aren't issued patents, they're published applications for patents.
When you read these applications, look first at the claims. If you believe any of those shouldn't be granted by the PTO in an issued patent, read the application in detail to learn more about the claim(s) you don't like. Why?
Because if you're right (example: claim shouldn't issue because of citeable prior art), you (anyone) can make a Preissuance Submission to the PTO saying why the claim(s) shouldn't be allowed. Here's a real brief view of how P.S works [1].
Really, if you as a person with domain knowledge see a problem with an application, it serves all of us for you to challenge it.
I've done this once in an energy storage context. The target application never issued as a patent (and rightfully so).
>Really, if you as a person with domain knowledge see a problem with an application, it serves all of us for you to challenge it.
If the PTO wants people to help them with review like this then the triple damages for having read a patent needs to be dropped (by congress). As it stands, corporate policy just about everywhere is that engineers shouldn't read patents.
The law around willful infringement (read: triple damages) changed recently.[1] It is perceived to be much more difficult to find willfulness now, and is generally determined by a jury. It is much more in line now with the common understanding that most people have, not some asinine bright-line test based on if you read a patent.
tl;dr: just reading a patent probably isn’t going to land you in triple damages by itself.
It is the idea that you shouldn't look for prior IP because knowledge of that prior IP makes you liable for triple damages because it's "willful infringement" rather than ignorance or accident. I have heard it before, I think it's ill-advised but I guess if you're 3M you have the legal team to deal with the fallout.
Regarding these particular patent applications (yes, even the broad claims and not just the titles and abstracts), I am baffled how willful infringement couldn’t apply.
I mean, you could just look at the papers the Google researchers themselves published, and just consider the sections on prior work and the bibliography.
Claims, even about a specific software system embodying an RNN device for a specific audio sample generation program, would obviously include research effort by Google staff to understand prior IP, much of which their neural network software mechanisms are inherently and directly based on in a way that most experts would view as clearly dependent on prior art.
What frustrates me is seeing so many people in the thread trying to defend Google and rationalize these applications as defensive patents, assuming the best of Google, and gainsaying people who are worried about the broad effects of the patents by obsessing over the mostly irrelevant fact that titles & abstracts don’t determine the content of an award.
There are a few other aspects to the problem that are also very serious.
- the popular belief that Google has some patent war chest in neural network devices can create a hugely chilling effect on other research, especially by other corporate research teams, borne purely out of beauracratic risk-aversion. Compliance and legal teams can and often do put the kibash on whole research programs just out of unnecessary risk aversion. In computer vision, just consider the effects of the patents granted for SIFT and SURF, which are many orders of magnitude smaller in scope than even the most graciously assumed narrowing of these Google patent applications. And SIFT and SURF had big chilling effects on various lines of research back when manual scale-invariant feature construction was state of the art. I even worked for a company that required my team to abandon and remove all use of OpenCV period once SIFT was added, just because we could, in theory, be sued into proving that we weren’t using SIFT through OpenCV. Companies will take these bonkers extreme stances. It’s bad for innovation overall. The corresponding effects for these Google applications would be huge: no TensorFlow, no Keras, unless it’s used via vendor lock-in with cloud providers that either are Google or pay Google — to be clear even if the ultimately awarded patents are much narrower in scope.
- The other huge thing is that Google puts a burden on civic infrastructure here. They are requiring the PTO to analyze the applications, and the public is trusting that the PTO has adequate expertise, time and budget to get the answer right. But Google knows that by making the claims overly broad, when they are inevitably walked back to narrower claims, the likelihood for PTO to get this completely right is super low, and with lobbying any of the round of error is sure to go in Google’s favor.
So even if you don’t see it as a bad faith action from the specific patent applications, it’s clearly bad faith in terms of inducing secondary effects of extreme legal risk aversion and basically wasting taxpayer resources to play this stupid cat and mouse game of how to narrow down the claims.
When I ask myself what is the long term self-interest here, it seems clear and in lock-step with other Google strategies: to try to own all the pipes by which modern machine learning can be commoditized, and to try to create a future where machine learning is only consumed as a service from the hybrid cloud-and-consulting platforms of a few entrenched incumbent corporations.
Maybe it made sense before the internet? I'm definitely not arguing that these patents are a good idea, and the argument that they're defensive feels weak since just straight up publishing puts it into the public domain where it acts as prior art to prevent other patents. The only way defensive patents make sense versus just publishing in the open is if you're using them as a threat against other people with patents in your area. The whole thing is gross.
Patents never made sense for software, but the horse has bolted and the law doesn't look like it's changing soon.
Defensive patents made sense until it turned out that most offensive action was from patent trolls who are not trying to create a product so they don't need to defend anything.
You're right; DeepMind has applied for patents covering this work, but the patents have not yet been granted. I meant to write "covered work" but after some editing ended up with "patented work" instead, which is inaccurate. Thank you for pointing it out.
These aren't even software patents. They are patents on mathematical procedures.
Just skim over US20180075343A1 [0] to see what I mean. It's not a patent on a particular application of sequence processing using neural networks. It's not a patent on a particular neural network architecture. It's a patent on the generic use of neural networks for sequence processing. It's not a specific process, it's an idea that encompasses a broad class of processes. There is no sane patent system in which that should be patentable.
Frankly, I don't have a problem with patenting specific algorithms or software implementations. Those, to me, are specific processes. These DeepMind patents are on another level.
Edit: I'm not a patent lawyer. Please correct me if I've read that patent wrong. It's basically a neural network mini-textbook/literature review.
Often a patent will include background information to assist the reader in understanding the claimed invention. The text will usually also point out the shortcomings of prior inventions and how the present invention is an improvement. Neither of those are part of the invention being claimed.
To understand the scope of what is claimed, you always need to look at the claims. The claims will always specifically point out the scope of the invention. Note that all elements in a claim are necessary. If an accused infringer's activity is missing an element, then it does not infringe.
I think that if you look at the claims of the patent you reference, you'll find that the claims are limited to particular applications of sequence processing using particular neural network configurations.
Not sure about the patentibility of those patents (they are still all in the application phase, and I have not read them), but to my understanding patents a) have to be tested before court and b) often fall back to less generic claims in practice (claim no. N, with N > 1), if not being completely invalidated. Now, I expect Google to have good patent lawyers, but time will show the value of those patents.
Usually when patent applications are made they are written up to be as general as possible and will then steadily be whittled down until they're accepted.
That said, I really don't see how a number of these will get even close to validation. Quite a few are describing an absolutely generic, everyday process with so much prior art it's insane (recurrent neural network?). If I'm missing anything here I'd love if someone could correct and inform me.
Many times companies will file patents on any IP they design that is broad enough to disrupt their market. With the patent being exceptionally broad, it ironically leaves less room for questions when the patent request is filed. One would hope everyone on the Government side of patents would be well versed enough in the given technology but unfortunately that is rarely the case so having a broad patent is sometimes easier to push through.
You are reading it wrong. The things they are patenting look like the following:
1. A neural network system implemented by one or more computers, wherein the neural network system is configured to generate an output sequence of audio data that comprises a respective audio sample at each of a plurality of time steps, and wherein the neural network system comprises:
- a convolutional subnetwork comprising one or more audio-processing convolutional neural network layers, wherein the convolutional subnetwork is configured to, for each of the plurality of time steps:
- receive a current sequence of audio data that comprises the respective audio sample at each time step that precedes the time step in the output sequence, and
- process the current sequence of audio data to generate an alternative representation for the time step; and
- an output layer, wherein the output layer is configured to, for each of the plurality of time steps:
- receive the alternative representation for the time step, and
- process the alternative representation for the time step to generate an output that defines a score distribution over a plurality of possible audio samples for the time step.
This appears to me to be describing an autoencoder for audio compression.
There are two other claims as well, all related specifically to audio processing.
Now whether those actual claims are valid is a separate question. I don't know the state of the art in 2016, they could have been. They are still math, I agree. But that's how you tell what they are actually claiming, you can entirely skip the body of the patent and just read the numbered claims.
The ones that don't start with "The ... of Claim ..., wherein" are independent claims meaning that they stand on their own. The ones that do read that way are typically used to narrow the independent claim so that when they say the independent claim is too broad they can make those dependent claims mandatory.
So for instance, if the USPTO (probably) rightly claims that generating audio with a convolutional neural network isn't patentable, they can fall back to saying "The neural network system of claim 1, wherein the audio-processing convolutional neural network layers include one or more dilated convolutional neural network layers." is valid combined with claim 1 because no one has thought of that specifically, and perhaps that's all that is granted.
In that case, as long as you don't use a dilated convolutional neural network layer then it is not infringing.
It's still all ridiculous nonsense, but that's how you read the patent.
Yes, the ratio is pretty low for DeepMind. Higher for Google Brain.
Still, DeepMind open-sourced two major frameworks: Sonnet (general DL) [0] and DeepMind Lab (RL environment) [1].
Frameworks != reproducible results. Ideally, they'd release all the code you need to get the results from the DDPG paper. I recognize that the code probably depends on proprietary DeepMind tools, making this not feasible. Still, it sucks.
Do you know that number? I have yet to hear a single instance or even rumor of that happening. Your post is the first time I've even seen it mentioned. Do you have any proof to back it up? When you make bold claims like that, I think the onus is on you to back it up.
I am not aware of them ever doing that. There is only two cases where there has been any IP aspect that I am aware of?
One started by Moto before them purchasing and the other the Waymo trade secrets case with Uber.
Google has purchased tons of patents but they use for cross license and NOT for royalities.
Are you confusing Google and Microsoft? MS does use their patents for royalities and perfect example is getting more per Android phone sold from royalities than Google gets.
They will sue if they see an emerging threat to the survivability of their core buisness. No one can touch Google on their ads and search ecosystem so they won’t sue. This is an illusion of benevolence.
There is two cases of protecting IP in the history of Google. The first started at Moto before Google purchased and they let continue.
The other was the best known which was the Waymo suit with Uber.
Google gets patents to protect themselves from someone else getting but do not use.
Google instead gives away tons of IP. Like Borg with K8s. Or VP8 and then on top indicates they will protect anyone that uses from patent infringement and on top gives for free.
Or SPDY or map/reduce or GFS or so many others could have patent and hurt the industry but instead gave away and helped the industry push forward. Can't think of any other company that has given away more valuable IP than Google.
Maybe. But also realize you have to protect patents. So if they do not which clearly they do not then no issue.
Also realize Google culture was there when they started and weak and now they are dominate it continues.
Perfect example is owning Chrome and Android which are the dominate two clients. Then own the top 2 web sites with Search and YouTube.
They create a better protocol which saves money with SPDY. Gives a better user experience. But even thought they have NOTHING to gain sharing they do and give away to the standards committee. They do NOT use as a competitive advantage which is just standard good business practices.
That is just not what MS did when in power.
Instead they give it away to help everyone. Also so much very valuable IP they just give away to the industry to help everyone. Kubernettes is a perfect example. But there is so many more. Why on earth do you help your competitors?
But they help competitors that do just evil things against them. They give away Android to Amazon who uses on their Echo, Dot and Fire among other hardware. In return Google gets a ban for ALL companies being allowed to sell the competing Google products. Not just Amazon will NOT sell as that is understandable but they will NOT allow anyone to sell them!
But then SPDY is changed and Google goes back and pays the expense of supporting the changes.
> But even thought they have NOTHING to gain sharing
This is wrong. They would have been ripped apart by people on the Internet if they switched to a proprietary protocol specific to Google for Google properties. Tech users brought the world to Chrome, tech users would just as quickly take them away if they pulled a Microsoft.
>Kubernettes is a perfect example.
K8s is not Borg. It was rewritten in open source. Xooglers could have started it just as easily without Google because the idea is very simple to spread.
>They give away Android to Amazon who uses on their Echo,
It's was open source when they got it and they capitalized on that in the marketing. They didn't "allow" amazon to do anything. They don't allow amazon access to the Google play store so they aren't exactly playing friendly.
>In return Google gets a ban for ALL companies being allowed to sell the competing Google products.
No, the ban is for not allowing native Google apps on Android users (a.k.a Fire phones) that don't pay the tax to Google. Tit-for-tat, which is petty, but tit-for-tat nonetheless.
> Not just Amazon will NOT sell as that is understandable but they will NOT allow anyone to sell them!
I bought a Chromecast at Walmart. Seems like Amazon is only preventing them from being listed in Amazon's marketplace.
>But then SPDY is changed and Google goes back and pays the expense of supporting the changes.
Every browser pays this. That's the Internet. Walled ecosystems with closed protocols die. Google chose the correct move by going open standards, there was no benevolence here.
Google only allowed encrypted traffic with SPDY and why is was used for so long and nobody even knew they were using. So nobody would even have known as nobody did know.
Between their own browser and OS and their own services do not think they would be torn apart. It is not like their services would only work on Chrome or anything like that just better.
Really no different than any browser does proprietary things to make theirs work as well as possible.
But it does not matter. Google gave SPDY away as wanted to push the industry forward and how we got http2 so fast as they basically handed to the standards group not only complete but also since been using for a while they even had the data also to support the benefits.
Why so many companies adopted so quickly. If not for Google we be still debating this and that and not be getting the benefits of http2.
On Borg. Yes K8s is a rewrite from what was learned not only from Borg but another Google solution whose name escapes me.
The point is Google spent a lot of money building Borg and learned from mistakes and fixed and the knowledge gained is extremely valuable yet Google gave it away with K8s. Now Amazon, Microsoft and most other cloud players have it for free and do not have to invest into going through what Google had to in gaining the knowledge. For free and no strings attached and use to compete against Google.
Why? Because it helps move the entire industry forward. Google just looks at the broader tech industry very differently than Microsoft.
Given HN's recent bashing of Google for abandoning it's Don't be Evil mantra this seems like a rather absurd position to take. "Google totes won't abuse this at all!"
It is complete and utter horse shit they are trying to patent things invented decades ago.
2018/081089, "Processing text sequences using neural networks", Priority date: 26 Oct 2016: "In one aspect, a system is configured to receive an input sequence of source embeddings representing a source sequence of words in a source natural language and to generate an output sequence of target embeddings representing a target sequence of words"
From the universal lack of a response on this board of anyone excited to read the patent applications to learn how to get AI to make sounds and such, clearly the US patent system is no longer serving that purpose.
Is it that bad in CS? In my field I'm constantly reading old patents for recipes to solve my problems. But I guess a lot of patents in chemicals are expired comparatively.
Remember that patents are not judged based on title alone! You can't even judge based on the claims in these publications because they are in the application phase, these are not issued patents. The claims that may be eventually granted will define the scope of protection afforded by these patent applications.
They could likely be defensive patents. In a system where patents are abused by trolls and used as negotiation leverages by corporate lawyers (I'm looking at you, Oracle and IBM), its better to file a patent for everything and never actually sue anyone for using those technologies.
Of course this is assuming that Google management does not abuse it themselves.
It's interesting to see Google's AI principles [0,1] published on the same day as these patent applications. Of course, this could be unrelated and complete coincidence. However, this _could_ have been strategic depending on the possible breadth* of what they are trying to claim. (And to be clear, I am not saying this is why they released these principles -- plenty of other reasons for that.)
*Note on the applications - I have not reviewed the claim sets so cannot comment specifically on what they are attempting to claim. In any event (and as others have noted, as well) the devil is in the details (e.g, the claims in each application -- more specifically, what is eventually allowed/granted).
> It's interesting to see Google's AI principles [0,1] published on the same day as these patent applications. Of course, this could be unrelated and complete coincidence. However, this _could_ have been strategic depending on the possible breadth of what they are trying to claim. (And to be clear, I am not saying this is why they released these principles -- plenty of other reasons for that.)*
It's almost certainly not a coincidence. Patent applications with possible international impact are typically filed before any publication, so as to avoid forfeiting any potential foreign patent rights. [0] Patent attorneys get extremely busy before, e.g., trade shows, new-product announcements, etc. — they want to get applications on file to preserve foreign rights.
Totally, you are 100% correct w.r.t. public disclosures and patent filings -- that you want to get the latter done before the former so as to preserve foreign rights outside of the US. Also, more broadly, so as to preserve any rights at all -- following the America Invents Act (AIA) the US is now in a first to file system (like the majority of the rest of the world), not a first to invent system [0,1].
In this case of Google that we are discussing, however, the patents have already been filed but not yet published -- thus, the public did not/could not know what the patents were. Patents publish some time after filing, often times 18mo after the filing date for the USPTO [1] but this timeline can vary for a number a reasons, e.g. an earlier priority date, etc [2].
That all said, much of the work here by Google has likely been disclosed already, but the existence of the patents and what was in them is generally new info for the public and industry.
The referenced AI principles post by Google discussed more about "values" and less about any given tech. That said, my original comment was aimed toward the possible notion of strategic PR (e.g., saying we are going to do 'good' with not just this tech, but also these patents which you are just learning about), not strategic IP (e.g., preserving defensibility for any specific tech).
> following the America Invents Act (AIA) the US is now in a first to file system
This is partly true. If Alice invents a widget and publishes details without filing a patent application, she still has a one-year grace period in which to file a U.S. patent application [1], and if she's successful she'll get only U.S. patent rights (plus a few other countries, I forget which). But whether or not Alice publishes before filing, the longer she waits, the greater the risk that Bob will independently invent the widget, without deriving it from Alice, and will file first. Even after the AIA, we still aren't a "first to file" system — we're a "first inventor to file" system.
Ah — I didn't look at the dates and assumed that the publication was the same date as the patent applications' filing, as opposed to the same date as their publications. In the latter case, it would indeed be a coincidence.
They're almost certainly not as broad as the title would suggest. You've got to look at the claims to figure out what the patent apps seek to cover. The title does not define the scope of activity they seek to cover. Usually, it just specifies the field in which they purport to have invented something.
Did you actually read the patent? Or did you just read the title? The title of the patent is not the patent itself. The details of what they are patenting is described in (sometime very longs and numerous) pages of technical description. I'm an engineer at a big tech company and I'm an author on many patents that have pretty generic/vague names. The actual invention is described in depth in the patent application and that's what's important
Assuming a patent issues with the claims as presented, all of the elements need to be present for there to be infringement. Claim 1 of the application you cited takes up almost an entire page of text and comprises more than a dozen limiting elements, of which you've paraphrased two. If there's a devil here, it's a lot smaller than you're implying it is.
They can't be, unless Google decides to enforce them. It's not the first time they crowd-source massive amounts of specialized effort, only to steal all the data, close them off and get away with it.
Panoramio is the biggest one I know. They essentially stole 10 years worth of effort from a community collecting carefully positioning millions of geo-tagged photos.
Then you have Youtube. It's not completely closed off, but you can't query the video database, you can't discover any of the old content, and they "delete" (hide from the public) all issues they consider controversial or against their political beliefs.
Now it's possible they used the good old 3Es to stimulate academic research relevant to their targets, use it to create useful products, and then patent everything to choke out the competition. Of course they have better ways to kill competitors (imagine the massive potential for undercutting they have, and how nobody cares about undercutting in the online world).
One of the interesting disconnects I see between layperson understanding of patents and legal understanding of patents is how one person's "broad and obvious" is another person's "novel and innovative," and the system does its best to handle that but the tension is forever there.
It's illustrative to remember that Thomas Jefferson himself was sued for using a patented Archimedes Screw design on his property, that he was extremely certain he could have fought in court and overturned since the 'novel' technology in question was known to the Greeks... and he paid the patent-holder anyway because he wanted to encourage patent adoption and use. The system has forever had this tension baked in, and we just muddle through generation after generation.
“This could be... potentially disastrous for innovation.”
This is how I feel about the patent system in general. I genuinely think parents harm innovation by setting artificial limits on innovation in a misguided attempt to incentivise innovation.
I hope they don't get the patents at all. You shouldn't be tied to DeepMind/Alphabet's goodwill over patents in fields they're clearly not the only ones innovating in.
Trump's recent USPTO appointed seems like a patent extremist. He recently argued that the USPTO (which he now leads) should not be allowed to call patents invalid after issuing them. He said the Supreme Court was wrong to rule that.
And I believe the previous USPTO chief was a former Google employee. So, either way, I think Google will get these, unless some organizations like the EFF start going through the process of attacking these patents.
Good thing it's the Supreme Court then, and not the "Maybe do what we say Court."
I get a little excited every time some idiot from the current administration tries to take on US rule of law, because I know the slow machinations of the legal system will eventually spank them into orbit over it.
As if the people of the United States haven't been signing blank checks to the executive branch for decades, and then just flooring the accelerator after 9/11. Your inability to get past the sideshow face of Trump's administration seems to have caused you to completely forget about, say, the free reign the administration has over its drone assassination program? They literally drop bombs on the phone numbers that Palantir spits out. Or National Security Letters that basically give them the legal machinery to completely and wholly subjugate an entity in any way they wish, with rubber-stamp "judicial approval" from a secret court and zero recourse for the victim, including requiring the victim to lie for the government to cover up its crime? Don't let the bullshit with your favorite color make you lose sight of the absolute power "the slow machinations of the legal system" has been corrupted to allow.
This is extremely concerning. Processing sequences using CNNs is particularly bad. It covers using CNNs to process ANY sequences, text, audio, video, etc.. Basically covering any potential use for a CNN. This would like a company patenting Welding. It's just so insanely broad.
I am not saying people shouldn't be outraged, but please read the actual claims rather than just reading the patent titles. Titles are just a general description of the invention. Inventor don't patent titles, they patent a claim that is described in a long technical patent application. DeepMind is not patenting the general idea of a "recurrent neural network", they just have a claim that can be described as recurrent neural network. In the actual patent documents, they need an actual in depth technical description of a new process or invention that is not covered by any existing publication.
Now do people see why it matters that it was Google the one building military drone AI, and not just "Random Company X" working with the Pentagon on this?
Google is far far ahead of others in AI, with or without the patents. The patents just ensure they also get to have a legal monopoly on some stuff to ensure that nobody else can catch-up to them in the technology area.
If your work was "described in a printed publication, or in public use, on sale, or otherwise available to the public" before the patent's priority data, it is available as prior art.[1] That doesn't mean you're immune, per se, but it does mean that nobody can patent exactly what you had made available.
If so, perhaps it would be within your interest to submit a Preissuance Submission showing prior art to the PTO, like ridgeguy mentioned in a sibling to your comment.
Patented work includes:
* 2018/048934, "Generating Audio using neural networks", Priority date: 6 Sep 2016
* 2018/048945, "Processing sequences using convolutional neural networks", Priority date: 6 Sep 2016
* 2018064591, "Generating video frames using neural networks", Priority date: 6 Sep 2016
* 2018071392, "Neural networks for selecting actions to be performed by a robotic agent", Priority date: 10 Oct 2016
* 2018/081089, "Processing text sequences using neural networks", Priority date: 26 Oct 2016
* 2018/083532, "Training action selection using neural networks", Priority date: 3 Nov 2016
* 2018/083667, "Reinforcement learning systems", Priority date: 4 Nov 2016
* 2018/083668, "Scene understanding and generation using neural networks", Priority date: 4 Nov 2016
* 2018/083669, "Recurrent neural networks", Priority date: 4 Nov 2016
* 2018083670, "Sequence transduction neural networks", Priority date: 4 Nov 2016
* 2018083671, "Reinforcement learning with auxiliary tasks", Priority date: 4 Nov 2016
* 2018/083672, "Environment navigation using reinforcement learning", Priority date: 4 Nov 2016
...
Is there anyone from DeepMind on HN who could comment on this?