In This Section
INNOVATE is the online magazine by and for AIPLA members from IP law students all the way through retired practitioners. Designed as an online publication, INNOVATE features magazine-like articles of 500-1,500 words in length on a wide variety of topics in IP law.
Publishing an article to INNOVATE is a great way for AIPLA members to build their brand by increasing recognition among peers and setting themselves apart as thought leaders in the IP industry.
Any current AIPLA member in good standing may submit an article for consideration in INNOVATE throughout the year. IP law students are especially encouraged to submit articles for publication.
Articles submitted to firstname.lastname@example.org are reviewed by an ad-hoc sub-committee of volunteers from AIPLA's Special Committee on Publications, the Fellows Committee, and other AIPLA peers. The review process ensures submitted articles follow the Guidelines for Article Submission. Approved articles are published quarterly.
Don’t miss your chance to be published with AIPLA’s INNOVATE! Email your article submission to email@example.com to be considered for the next edition.
Patenting Machine Learning Inventions for Companies Outside the Software Industry
By Gregory Rabin
The use and development of artificial intelligence and machine learning technology is spreading from niche Silicon Valley software companies to almost all innovative businesses, including those outside the traditional software field. Stitch Fix CEO Katrina Lake recently said: “In 10 years, every ‘relevant’ company will be a tech company.” As such, a strategy for inventing and using machine learning technology, and for generating IP based on the inventions, is essential to non-software companies.
While some large software companies have been patenting artificial intelligence-related technologies for over a decade, non-software companies might be new to patenting these types of inventions. This article discusses how these technologies might be used or patented by companies outside the software industry. This article is also intended to assist private patent practitioners, as well as in-house attorneys, in understanding the machine learning technology space, developing processes for identifying patentable inventions in the machine learning field, and patenting such inventions.
Uses of Machine Learning Technology by Companies Outside the Software Industry
One example of a company outside the software industry that uses machine learning technology is Atomwise, a biochemical discovery startup in San Francisco. Atomwise deployed a deep learning machine, AtomNet, to tackle key real-world issues in improving pesticides. Deep learning allowed Atomwise researchers to simulate millions of compound and identify the ones that target pests without causing toxicity in humans or other friendly species. Using traditional research methods, simulating millions of compound would be impractical. This approach has allowed Atomwise to produce less harmful pesticide products faster than its competitors.
One big benefit of machine learning technology is faster iteration. Samples can be generated and tested much more quickly by machines than by humans. Dilbert creator Scott Adams tried many different cartoons before he developed Dilbert and became successful through it. The iteration process took Adams many years. However, a machine learning algorithm that generates cartoons can do this much more quickly. Similarly, in the technology space, faster iterations allowed Atomwise to develop pesticides much more quickly than its competitors who used traditional schemes.
Not surprisingly, there is a strong upward trend in patent application filings for machine learning technology, which will likely continue into the future. Patentable inventions may include: (1) new approaches using machine learning technology to solve problems in the company’s line of business or (2) new machine learning technologies themselves.
New Approaches to Problem Solving
The patentability of new approaches to problem solving in a non-software company’s line of business may be based on Diamond v. Diehr. At issue in Diamond v. Diehr was the validity of U.S. Patent No. 4,344,142, titled “Direct digital control of rubber molding presses,” and issued to James R. Diehr, II, of Troy, Michigan. The Supreme Court held that controlling the execution of a physical process, by running a computer program did not preclude patentability of the invention as a whole. The novelty of the invention could lie either in the computer program or in the physical process itself.
Importantly, the Diehr patent is not a machine-learning invention. In fact, it relates to “ancient” computer technology – having been filed in 1975 and issued in 1982. This invention is not machine learning. Rather, it leverages a simple 1970s computer being preprogrammed with instructions. In the Diehr patent, the computer gathers data from the data storage and the environment (mold temperature) and makes decisions based on this data using preprogrammed rules. This was sufficient for a patentable invention according to the court in Diamond v. Diehr. Replace preprogrammed rules with a trained neural network (and a description of the training process), might yield have a machine learning invention that is clearly patentable under the rules expressed by the Supreme Court in Diamond v. Diehr.
Examples of technologies that could be patent-eligible based on the Diamond v. Diehr precedent include using a trained neural network to control fabrication of a chemical (e.g., rubber, glass, detergent, etc.) or using a trained neural network to control development of a biological or biomedical compound (e.g., a drug, a vaccine, an artificial limb, an artificial organ, lab-grown meat, etc.). These inventions may have easily-resolvable (if any) hurdles under 35 U.S.C. § 101.
One modern example of this approach is U.S. Patent No. 8,478,535, issued to Nebojsa Jojic of Redmond, WA, and originally assigned to Microsoft Corporation. This patent application was filed on December 30, 2005, and issued on July 2, 2013. It is titled, “Systems and Methods That Utilize Machine Learning Algorithms to Facilitate Assembly of AIDS Vaccine Cocktails.” In summary, machine learning techniques are used “to generate vaccine cocktails for species of pathogens that evolve quickly under immune pressure of the host.”
New Machine Learning Technologies
Non-software companies may seek patents for new machine learning technologies that their employees invent. Different problems may require different machine learning technologies to solve them. Typically, when faced with a new machine learning problem, software companies attempt to recycle preexisting technologies to solve them. For example, if the problem is cat facial recognition, the solution could be re-training a human facial recognition model on cat faces. However, this might not always work. For example, replacing human faces with cat faces might work. However, replacing human faces with complex chemicals might be more difficult. Changes that could be made may include one or more of: (1) a new feature vector or new data studied by the neural network, new training datasets, and a new neural network structure. For new machine learning technologies, patentability may be based on advances in the field of computer science itself, rather than the use of a computer to solve a problem in another field. For example, the technology may relate to a new neural network, rather than using a previously-existing neural network to control a process of curing rubber.
It should be noted that some technologies, such as primarily mathematical algorithms or non-technological business algorithms that do not impact anything outside the computer, might not be patent-eligible. Examples of such technologies include techniques for hedging investment risk, and computations done inside a computer that do not have a real-world impact and are not tied to the functionality of the computer. Examples of things that are tied to the functionality of the computer include a central processing unit (CPU), a graphics processing unit (GPU), a memory structure, a data structure, and the like). For these technologies, the inventors’ companies may wish to consider a defensive publication to prevent competitors from getting a patent if the law changes.
Drafting Patent Applications
The specification should apply the claimed technology to a real-world use case. To overcome potential rejections, the specification should disclose as many technical details as possible about inventive technology. The specification should explain not only which computations are done, but how they are done, in as much granularity as possible. For example, the specification may focus on which data structures and hardware elements are used and how the data structures or hardware elements interact with one another to perform the invention. The specification should illustrate why the invention is better than previously existing solutions or other solutions that a competitor could invent. The specification should show that the invention is a technical solution to a problem that is technical in nature.
Claims should be drafted with infringement detection in mind. For example, if possible, the claim should clearly state the input and the output of any algorithm and, to the degree possible, avoid focusing on the “black box” of the internal workings of the technology.
Validity Under 35 U.S.C. § 101
In issuing (or accepting arguments to overcome) § 101 rejections, Examiners typically want to see how the technology is tied to a practical application and is “real” and not “abstract.” Is the invention simply doing calculations within a computer or do those calculations have a real world impact outside the computer. Non-software companies, which develop machine learning technologies for real-world industrial applications, rather than for some potential not yet defined applications, have a strong advantage here. The USPTO guidance of January 7, 2019, takes the position that:
“If a claim recites a judicial exception (a law of nature, a natural phenomenon, or an abstract idea…), it must then be analyzed to determine whether the recited judicial exception is integrated into a practical application of that exception. A claim is not ‘directed to’ a judicial exception, and thus is patent eligible, if the claim as a whole integrates the recited judicial exception into a practical application of that exception. A claim that integrates a judicial exception into a practical application will apply, rely on, or use the judicial exception in a manner that imposes a meaningful limit on the judicial exception, such that the claim is more than a drafting effort designed to monopolize the judicial exception.”
However, it should be noted that courts have not given deference to the USPTO guidance. In responding to an office action, an attorney should argue for § 101 eligibility under both the USPTO guidance and any relevant case law.
The most relevant case law is the Alice/Mayo framework that set forth a two-step test, as supported by the Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014) and Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012) cases. In step 1, one determines whether the claims are directed to a patent-ineligible concept, i.e., an abstract idea. In step 2, one considers the elements of each claim (both individually and as an ordered combination), and determines whether the additional elements transform the nature of the claim into a patent eligible application.
In Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir. 2018), the Federal Circuit Court focused on step 2 of the Alice/Mayo framework to determine that, while Berkheimer’s independent claim was not patent-eligible, some of the dependent claims were patent-eligible. Berkheimer argued that the claimed combination improves computer functionality, in a digital asset management system, through the elimination of redundancy and the one-to-many editing feature, which provides inventive concepts. The court specifically noted the following patent-eligible improvements in the dependent claims: (i) eliminating redundancy, and (ii) a one-to-many editing process where a singular linked object, common to many documents or files, can be edited once and have the consequence of the editing process propagate through all of the linked documents and files (reducing the effort needed to update files).
As artificial intelligence and machine learning technologies become more and more ubiquitous, innovative companies outside the software industry are well advised to begin seeking patents in this space. Selecting knowledgeable and experienced artificial intelligence patent counsel may be beneficial in ensuring successful patent portfolio development.
 Eric Johnson, In 10 years, every “relevant” company will be a tech company, Stitch Fix CEO Katrina Lake says, Vox (Jul. 24, 2019, 6:20 AM), https://www.vox.com/recode/2019/7/24/20707751/katrina-lake-stitch-fix-retail-fashion-clothing-data-kara-swisher-recode-decode-podcast-interview.
 Tom Simonite, Despite Pledging Openness, Companies Rush to Patent AI Tech, Wired (Jul. 31, 2018, 7:00 AM), https://www.wired.com/story/despite-pledging-openness-companies-rush-to-patent-ai-tech/.
 Diamond v. Diehr, 450 U.S. 175 (1981).
 U.S. Patent No. 8,478,535.
 Federal Register Vol. 84, No. 4, Jan. 7, 2019 (available at: https://www.federalregister.gov/documents/2019/01/07/2018-28282/2019-revised-patent-subject-matter-eligibility-guidance).
Greg is a senior patent attorney at Schwegman, Lundberg & Woessner. He has been practicing patent law for a decade. Prior to joining Schwegman, Greg practiced at McDermott, Will & Emery in Boston, MA. Greg holds a J.D. from the University of Michigan Law School, dual Bachelor’s Degrees in Computer Science and Mathematics from MIT, and a Master’s Degree in Computer Science from MIT.
Greg is the lead author and editor in chief of Bloomberg BNA’s electronic book about patenting artificial intelligence inventions: “Artificial Intelligence and Machine Learning – Protecting the Next Ubiquitous Technology.” (Publication expected in 2020.) The book covers considerations when patenting artificial intelligence technology in the United States and abroad, and legal issues with technologies that are conceived or reduced to practice using artificial intelligence.
Greg has spoken about patenting inventions in artificial intelligence and machine learning before the American Intellectual Property Law Association (AIPLA), the United States Patent & Trademark Office (USPTO), and several continuing legal education (CLE) providers. Specifically, Greg has discussed (i) overcoming 35 U.S.C. § 101 (patent-eligible subject matter) rejections for artificial intelligence inventions, and (ii) patenting artificial intelligence inventions for companies outside the software industry at various AIPLA events. In addition to machine learning and artificial intelligence, Greg has also drafted and prosecuted patents related to mobile and WiFi networks, operating systems, cryptography, security systems, control systems, and robotics.
Greg has drafted and prosecuted multiple patent applications to issue in the United States and abroad. He has worked with European, Chinese, Japanese, Korean, Taiwanese, Indian, Canadian, and Australian counsel to prepare and prosecute foreign patent applications for his clients. Greg is a native speaker of both English and Russian. Greg conducts “patent mining” sessions with clients, where he visits the client’s office, meets with inventors and in-house counsel, and identifies inventions for patenting (or for coverage as a trade secret). Greg is very communicative and believes in keeping clients abreast of what is happening with the patent applications that are entrusted to him. Greg frequently leverages Patent Attorney-Examiner Interviews, and tries to conduct an interview before submitting a response to a US Patent Office Action. Furthermore, he regularly takes advantage of the Patent Office’s After Final Consideration Pilot (AFCP) program to interview Examiners in order to figure out how to bring the patent application to allowance or advance patent prosecution.