Request for Comments on Patenting Artificial Intelligence Inventions
federalregister.gov
federalregister.gov
Computer scientists are only eligible if they have an undergrad degree from an ABET-accredited program. This rules out people coming from AI powerhouses like CMU and Stanford, as well as Princeton, Yale, etc. Masters and doctoral degrees also don’t count, which is baffling since a PhD involves reading a huge chunk of the literature. The whole process is centered around 1960s “mechanical engineering as the source of all innovation” ideas and seems like it needs a massive update.
One is by having sufficient coursework in various categories, even if you degree itself was not one they recognize. There are four options if you do it this way.
They are 24 semester hours in physics; or 32 semester hours that includes 8 hours of either physics or chemistry and 24 hours in biology, botany, microbiology, or molecular biology; or 30 hours in chemistry; or 8 hours of either chemistry or physics and 32 hours in chemistry, physics, biology, botany, microbiology, molecular biology, or engineeing.
Under that last option, computer science courses count.
The other route in is by taking and passing the Fundamentals of Engineering (FE) test.
I was once seriously considering becoming a patent agent, but my degree is in math (Caltech, class of '82) which is not on the PTO's list of acceptable degrees, and so looked into these other routes.
The main issue with the coursework route for me was that when I was there Caltech graded all first year courses pass/fail, and the PTO does not count courses taken pass/fail. That knocked a year of physics, a term of physics lab, a year of chemistry, and a year of chemistry lab out of the running.
The FE test route looked quite reasonable. This was a long time ago, and I believe they have changed the format of the test since then, but when I was looking it looked like it would have taken maybe a month or two of spare time prep.
Still, this seems like a massive, unnecessary effort to establish that your Caltech degree, or my Yale one, are at least as good as one from DeVry Tech.
One way to attack this would be if researchers outside of Google and Facebook started patenting their discoveries, but then immediately released all rights to it under a license that's valid for anyone who does not sue offensively. That way if Google wanted to sue for infringing use of invention A, they would themselves immediately be infringing on invention B and C. Quickly any large player would be entangled into a web where any offensive lawsuit is impossible. If we had standardized licenses with these clauses, we could stand a chance to stop this.
If BatchNorm, ConvNets, LSTMs, Dropout, ReLUs or any such essential neural net component came with such a clause, I could see the big players accepting the license, especially since it aligns with their publicly stated claim that they only file patents for defensive purposes.
From the article:
> 1. Inventions that utilize AI, as well as inventions that are developed by AI
> 3. Do current patent laws and regulations regarding inventorship need to be revised to take into account inventions where an entity or entities other than a natural person contributed to the conception of an invention?
[0] https://www.eff.org/deeplinks/2014/06/bad-day-bad-patents-su...
An invention submitted by a group containing an AI co-inventor or using an uncredited AI as a tool could be in any field, not necessarily a software patent, so the "idea but implemented on a computer" doesn't really come into play here.
If the USPTO has trouble understanding a technology, they should err on the side of caution and NOT grant the patents, instead of asking for comments ex post-facto
Take for example Dropout. This is a ridiculous thing to patent. Dropout is essentially half your network malfunctioning. A computer that randomly caught fire could "come up" with the same "algorithm". How can the USPTO judge that patenting this thing in any way would help innovation (that's the purpose of patents). In fact, most researchers will now be discouraged from researching dropout-like techniques any further
For all the faults of patents (and there are many), one of the great benefits is that when you patent something, you must publish it for all to see. If we want to keep innovation moving along quickly, there needs to be incentives to publish your findings for everyone to review and build on.
The lawsuits don't really start until someone starts making big money. Yes, they are burdensome and expensive, but I'd rather keep innovation moving quickly and having the winner pay a tax than not having any innovation at all.
Do people tend to learn things from patent filings, or from other sources?
I've looked at a few patents, and they don't read like anything that tries to educate the reader.
When you file for a patent, it usually does not get published to the public for 18 months. In those 18 months, the inventors usually publicly announce themselves (they have patent protection, so there's little harm doing so). If they didn't have patent protection, there would be much more resistance to publishing your work.
There is more than plenty of open research in NNs. There is absolutely zero insights in these patents that is not already published.
I also find it a very weak argument in favor of patents. You can reverse-engineer anything, no need to read the filings
The general procedure for many about to publish a research paper in a growing field is to get a patent application on file first. There wouldn't be nearly as many publications if there were no patent protection.
It wouldn't be as open if there were no patents. When researchers working for a big company are about to file a research paper, their corporate sponsors get the patent application containing their work on file first.
This is so incredibly wrongheaded I don't even know where to begin.
Here's their patent: https://patents.google.com/patent/US5052043A/en
These patents are overly broad, overly premature. They should never have been granted in the first place.
I think the NN pioneers who are first names in these patents should set up a nonprofit themselves. Hopefully many of them still seem to believe in the ethic of science , having lived decades in which nobody cared about NNs
The current state is sort of a patent cold war where both sides have weapons but they don't use them. So we get the advantages of patents (disclosure) without many of the potential disadvantages.
It’s also not as broad as patenting backdrop.
Also, the AI/ML parts of inventions I have seen are based on conventional/non-patentable AI/ML techniques.
It is the combination of AI/ML components with other components, such as, the bits about managing the inputs, outputs, feedback loops, model selection/management, training optimizations, or the like, that make a patentable AI/ML invention. Further, IMO, it would be poor patent drafting to draft a patent application that relies on particular AI/ML techniques. If your AI/ML techniques are really novel and patentable, you should still draft the patent such that other AI/ML techniques could be substituted into the system as well as claiming the unique AI/ML technique.
4. Should an entity or entities other than a natural person, or company to which a natural person assigns an invention, be able to own a patent on the AI invention? For example: Should a company who trains the artificial intelligence process that creates the invention be able to be an owner?
Well, this seems a little dangerous. I would argue that any invention or innovation generated by an AI should be made public domain.
As we rapidly approach the possibility of genuine AI, the gap between the haves and have nots will increasingly be defined not by accumulation of capital but by accumulation and control of information. If the explosion of technical progress we've seen in ML recently continues, it's quite likely that future designs and breakthroughs will eventually come from neural nets themselves - and if we define these innovations as IP and afford the usual legal protections to the nets that generated them, as the question seems to imply ("other than natural persons"), then I imagine by proxy the ultimate owner of the patent is the owner of the net. Which forms the foundation of a dystopia defined by unprecedented "wealth" inequality where one or a handful of first movers become irreversibly entrenched as the gap between AI powered innovation and human powered innovation will widen exponentially once that door is unlocked.
I think much of the progress in the ML explosion is owed to the beauty of open source and open access publishing on arxiv, and I can't help but feel like getting neural network designs mixed up with patent law would stymie the iterative collaboration that defines ML research.