Considerations for Artificial Intelligence Patenting and Enforcement

The advancement of artificial intelligence technology raises important considerations of patent eligibility, inventorship and enforcement. Developments in the law, such as what constitutes patentable subject matter, along with the rapid progression of the technology, present real challenges for the protection of AI innovation. Although U.S. courts have provided some guideposts to navigate the relevant patent law issues, few have addressed AI patents directly, and much uncertainty remains for protection and assertion of AI-related intellectual property.

AI Technology

AI generally refers to the mimicking of human cognitive functions in machines. One example is an expert system, typically a computer system that simulates human decision-making to solve problems in a specialized area by applying a logic or rules-based model to a knowledge base. AI encompasses a wide array of tools applied to automated problem solving in applications ranging from computer vision and image processing, to speech recognition and natural language processing, to robotics and autonomous vehicles, to medical records and healthcare diagnostics.

More recently, a focus of AI has been machine learning. Generally speaking, machine learning attempts to mimic the human brain’s ability to draw conclusions or make predictions based on incomplete or imprecise data. The machine learns how to solve problems and complete tasks, often without requiring specific programming instructions. For instance, if a machine is learning to identify dogs from digital images, over time it may learn from mistakes, like incorrectly identifying a cat as a dog, and correct itself by modifying its algorithm.

Deep learning is a class of machine learning based on unsupervised learning of multiple levels of features or representations of data. It attempts to mimic the brain with simulation of layers of neural networks. Once a machine has learned how to perform a task in one set of data, it may then apply its knowledge to other data sets without human supervision. A deep learning algorithm could, for example, examine large data sets, such as shape, size, pattern, color and texture of objects in an image, rank these features, and identify animals with more specificity, such as specific dog breeds.

Patent Eligibility

In recent years, the U.S. Court of Appeals for the Federal Circuit has elaborated on the now familiar Alice/Mayo two-step framework for determining what is patentable subject matter under 35 U.S.C. §101. The decisions addressing software-related inventions offer some insight into the patent eligibility of AI inventions.

Step one of the test asks whether the claim is directed to patent-ineligible subject matter, such as a law of nature, natural phenomena, or abstract idea. If the answer is yes, the question in step two is whether any additional claim elements, either individually or as an ordered combination, constitute an inventive concept that transforms the claim into patent-eligible subject matter.

Since AI technology involves the replication of human brain functions, it can potentially raise concerns that AI inventions are ineligible mental processes under step one. On the other hand, AI at its core involves an improvement in the ordinary or routine operation of a computer – that is, to make the machine more human-like.

The Federal Circuit has emphasized the significance of improvements in computer technology in the step one analysis. In Visual Memory LLC v. NVIDIA Corp., the Federal Circuit ruled that claims directed to an improved computer memory system were eligible and not abstract. The claims expressly state that the invention’s improved memory system is achieved by configuring a programmable operational characteristic of a cache memory based on the type of processor connected to the memory system. In an earlier Federal Circuit decision, Enfish LLC v. Microsoft Corp., claims directed to a self-referential table for improving a computer database did not constitute an abstract idea. In McRO Inc. v. Bandai Namco Games America Inc., the Federal Circuit found a method of automatically producing realistic animated lip synchronization was “directed to a patentable, technological improvement over the existing, manual 3-D animation techniques,” not a patent ineligible abstract idea.

The Federal Circuit has acknowledged the potential overlap between steps one and two and that improvements in computer functionality are relevant in step two as well. For example, in Amdocs (Israel) Ltd. v. Openet Telecom Inc., the court found an inventive concept in the claim’s “unconventional technological solution (enhancing data in a distributed fashion) to a technological problem (massive record flows which previously required massive databases).”

The Federal Circuit has also distinguished technological improvements from abstract ideas lacking an inventive concept. In RecogniCorp LLC v. Nintendo Co., the patentee claimed the abstract idea of encoding and decoding composite facial images using a mathematical formula. The Federal Circuit found no improvement in the functioning of a computer. Indeed, as the court noted in step two of the test, the claimed method “does not even require a computer; the invention can be practiced verbally or with a telephone.” More recently, in Smart Systems Innovations, LLC v. Chicago Transit Authority, claims to a system and method for regulating entry in a transit system using information from a bankcard were found ineligible. Again, the Federal Circuit explained that there was no improvement in computer technology, but rather the mere invocation of computers in the collection and arrangement of data.

Notably, in his Smart Systems dissent, Judge Richard Linn remarked on the “great uncertainty” in the application of the Alice/Mayo framework and cautioned against “engaging in an overly reductionist exercise to find the abstract idea underlying virtually every claim.” According to Linn, this uncertainty is particularly troubling in certain fields including computing and AI.

Improving Computer Technology

Based on this guidance, AI software may be patent eligible on the ground that it involves improvements in the operation or performance of a computer. A key for claim drafters is to focus on the specific improvements in computer capabilities or specialized computer components. For example, if a system relies on deep learning to examine large data sets and perform tasks based on its own insights without human supervision, the claims should identify particularized steps of the deep learning method, such as the application of specific algorithms to data sets.

An illustration is U.S. Patent No. 9,552,548 (the ’548 patent) entitled “Using Classified Text And Deep Learning Algorithms To Identify Risk And Provide Early Warning.” Independent claim 1 recites training and re-training deep learning algorithms using risk classification datasets, and dependent claims recite particular types of deep learning algorithms, such as recurrent neural networks in claim 3. Claim 1 reads:

1. A method of using classified text and deep learning algorithms to identify risk and provide early warning comprising:

creating one or more training datasets for textual data corresponding to one or more risk classifications, wherein said risk classification comprises one or more threats or risks of interest;

training one or more deep learning algorithms using said one or more training datasets;

. . .

determining if said identified one of said one or more threats or risks of interest is a false positive or a true positive;

re-training said one or more deep learning algorithms if said identified one of said one or more threats or risks of interest is a false positive;

. . . .

The ‘548 patent, filed September 27, 2016 and issued January 24, 2017, recites only method claims. All application claims issued without amendment in an expedited examination.

Another example is Patent No. 9,547,821 (the ’821 patent), entitled “Deep Learning For Algorithm Portfolios.” The patent discloses generally an automated methodology for choosing the most appropriate algorithm (problem solver) for the problem at hand based on a descriptive set of features that represent the problem instance. Claim 1 recites:

1. A computer-implemented method of automated feature construction for algorithm portfolios in machine learning, comprising:

receiving, by one or more processors, a problem instance represented as text describing a problem to be solved by computer-implemented problem solver;

generating, by one or more of the processors, a gray scale image from the text . . . ;

rescaling, by one or more of the processors, the gray scale image . . . , the rescaled gray scale image representing features of the problem instance;

inputting, by one or more of the processors, the rescaled gray scale image as features to a machine learning-based convolutional neural network; and

training based on the rescaled gray scale image, by one or more of the processors, the machine learning-based convolutional neural network to learn to automatically determine one or more problem solvers from a portfolio of problem solvers suited for solving the problem instance, . . . .

The ’821 patent, filed February 4, 2016 and issued January 17, 2017, also includes system and computer readable storage medium claims. All application claims issued without amendment following a Track One examination.

For AI software, as with other computer-related inventions, claiming the technological improvement with specificity is key to patent eligibility. In Vehicle Intelligence & Safety LLC v. Mercedes-Benz USA LLC, the Federal Circuit found patent ineligible an “expert system” that provided faster and more accurate testing for driver impairment because neither the claims nor the specification provided any details as to how the “expert system” worked.

To the extent AI technology is implemented using specialized hardware, or other particularized elements of a machine, reciting such hardware or machine elements in the claims will tend to support eligibility. For instance, AI in an autonomous vehicle may rely on sensors as well as specialized chips or graphics processing units in addition to standard processors. Patent-eligible claims might incorporate such sensors and specialized processors.

Inventorship

Given that AI involves the simulation of human decision-making by machines, questions can arise regarding the inventorship of these machine processes. For example, if machine learning involves a computer modifying its algorithms based on a training dataset, is the inventor machine or human?

35 U.S.C. § 100(f) defines an “inventor” as the “individual or, if a joint invention, the individuals collectively who invented or discovered the subject matter of the invention.” In University of Utah v. Max-Planck-Gesellschaft, the Federal Circuit held conception is a mental act requiring performance by “natural persons.” Although this holding addressed conception by humans versus corporations or sovereigns, not humans versus machines, as a matter of policy, it seems unlikely that machines will ever be considered inventors.

Assuming that an inventor must be human, another question arises when a computer modifies its algorithms based on a training dataset: is the invention the training dataset, the algorithm, or both? In other words, does machine modification to an algorithm, as opposed to a human modification, detract from its patentability? Under Section 103, “[p]atentability shall not be negated by the manner in which the invention was made.” However, the Federal Circuit explained in Board of Education of Florida State University v. American Bioscience Inc. that teaching skills or general methods that facilitate a later invention, by itself, is not sufficient for inventorship.

Patent applicants should consider clearly explaining the connection between human involvement and subsequent algorithm modification, such as defining the problem for the machine to solve, reviewing and analyzing the machine’s operation, and determining whether the machine performed in a manner expected by the inventor. This becomes even more complicated with deep learning that involves feature or rule extraction with minimal intervention by humans. Such AI applications can use data in unique ways that are not necessarily a foreseeable result of its base algorithms.

Challenges also may arise where inventions incorporating AI are developed by multiple persons, such as a multidisciplinary team of computer scientists and mathematicians who understand an AI hedge fund algorithm, and the hedge fund managers and investment advisers who understand the industry sector data and investment strategies used to train the algorithm. Inventorship may depend on whether the invention is considered the algorithm itself, application of the algorithm to particular data sets, or other elements of the AI system. Potential conflict also arises where, for example, the data set is owned by a third party and licensed to the AI development team or end user.

While questions of joint inventorship are not unique to AI, the potential significance of independently developed elements, such as training datasets, to the patentability of AI inventions poses new challenges for the inventorship determination.

Enforcement

Like inventorship, infringement of AI patents presents nuanced factual scenarios that have not been directly addressed by the courts.

One enforcement consideration is whether patented AI methods are externally verifiable by detecting the algorithm or training data used in an AI system. Claim specificity is important here too in light of the autonomy of AI machines. For example, would different machines modify an algorithm in the same way, or at least within the scope of the claims, if presented with the same algorithm and data set? If the claim is broad enough, it may provide leeway to account for any differences between the autonomous behavior of machines, but broader claims are more likely to suffer patentability deficiencies such as ineligibility, as discussed above, and inadequate disclosure or written description under 35 U.S.C. § 112.

Another consideration is whether divided infringement can occur between an autonomous machine and another entity. Under the Federal Circuit’s holding in Akamai Technologies Inc. v. Limelight Networks Inc., one entity can be held responsible for another’s performance of method steps if the first entity directs or controls that performance. Who directs or controls a machine that is capable of learning without supervision? Is it the developer of the AI system, the provider of the training data or the end user?

Even assuming that the end user exercises direct control over the machine by invoking the AI system, there may be an argument for at least inducement by the developer or provider of the training data. Again, the answer will likely depend on the scope of the claims, and whether they focus on the algorithm, data set or their application.

Conclusion

The complexities of AI technology create challenges for obtaining and enforcing AI patents. Because AI at its core involves enhancing the functionality of a computer or machine, AI inventions would appear to easily satisfy at least step one of the Alice/Mayo analysis. However, existing jurisprudence offers no bright-line rules for the patentability of software-related inventions. Thus, patentees must be mindful to claim with specificity the particularized steps of an AI method or elements of AI-related hardware. And patentees may have to resist the temptation to seek broad claims, even in this relatively new field of technology.

The difficulty of inventorship and infringement determinations for AI inventions also counsel those in the field to proceed with caution– particularly when multiple parties are developing or using the AI technology. As we await further guidance from the courts on relevant patent law issues, AI innovators may consider alternatives to patent protection, such as copyright and trade secret protection.