In October 2017, Hyper Search brought a patent infringement action against Facebook in the District of Delaware, asserting U.S. Patent Nos. 6,085,219, 6,271,840, and 6,792,412. Facebook sought to dismiss the complaint under Rule 12(b)(6), alleging that all asserted claims were invalid under 35 U.S.C. § 101. While all three patents suffered a similar fate, the '412 patent and the Court's reasoning regarding this patent is of particular interest because the claims thereof specifically address machine learning.
Machine learning is a subset of artificial intelligence in which a data set -- usually a large one -- is used to train a computational model. This model is then applied to new instances of data to provide a result in line with the training. In this fashion, software can be used to classify and/or predict outcomes often with greater accuracy than can be expected from human performance or programs developed in a more traditional manner. Notably, the exact algorithm and rules used by the model are not explicitly provided by the programmer -- instead, the process of training the model develops these rules from the training data. Machine learning has been applied with success to various technological fields, such as speech recognition, image classification, supply chain management, drug efficacy, genetic similarity, and information technology services, just to name a few.
The patent eligibility of artificial intelligence, machine learning in particular, is an open issue. There have been no substantive Supreme Court or Federal Circuit § 101 cases to date, and just a handful in the district courts. But this handful has not been representative of how machine learning claims are now drafted (or should be drafted), and the machine learning content therein was often tangentially related to the actual claim language. Thus, any proceeding that can provide insight on the nuances of how machine learning inventions are being viewed by the courts is potentially helpful.
Why is patent-eligibility an issue for machine learning inventions? This is because the validity of software patents as a whole has been subject to an additional layer of scrutiny for the last five years.
In Alice Corp. v. CLS Bank Int'l, the Supreme Court set forth a two-part test to determine whether claims are directed to patent-eligible subject matter under § 101. One must first decide whether the claim at hand is directed to a judicially-excluded law of nature, a natural phenomenon, or an abstract idea. If so, then one must further decide whether any element or combination of elements in the claim is sufficient to ensure that the claim amounts to significantly more than the judicial exclusion. But elements that are well-understood, routine, and conventional will not lift the claim over the § 101 hurdle.
Considering machine learning from one viewpoint, it is no more than abstract mathematics enabled by generic computing devices and applied to data; therefore, it is inherently ineligible under § 101. From the other end of the spectrum, machine learning can be thought of as solving complex problems by creating new algorithms that in practice can only be used by computers; thus, it is inherently technical. Most uses of machine learning lie between these two extremes. Applying a vanilla model to any data -- even new data -- is probably not eligible and may even be obvious as well. On the other hand, most practical uses of machine learning require at least some innovative model design, training techniques, data preparation, mapping of the prepared data to the model, and post-processing of model output. Any of one or more of these latter mechanisms may be enough to lift a claim over the § 101 hurdle.
In any event, claim language and the problem being solved matters in the § 101 inquiry, as Hyper Search found out the hard way. Claim 1 of the '412 patent recites:
A system for controlling information output based on user feedback about the information comprising:
a plurality of information sources providing information;
at least one neural network module that selects one or more of a plurality of objects to receive information from the plurality of information sources based at least in part on a plurality of inputs and a plurality of weight values;
at least one server, associated with the neural network module, that provides one or more of the objects to one or more recipients;
the recipients enabling for one or more users to generate feedback about the information; and
wherein the neural network module generates a rating value for a plurality of the objects at the end of an epoch, redetermines the weight values using the rating values, and selects which objects to receive information during a subsequent epoch using the redetermined weight values and the inputs for that subsequent epoch.
In a nutshell, the invention uses a neural network (which is a type of machine learning model) to determine how to provide information from particular sources to particular output objects. The neural network periodically re-evaluates the weights based on ratings from the objects.
Facebook contended that this claim "is directed to the abstract idea of providing information based on feedback from recipients." Hyper Search countered that its invention is "an improvement in computer capability because the system controls information output based on user feedback about the information itself" and "that this is a specific process that is new, performed by a computer, iterative, and adaptive."
The Court appeared to adopt Facebook's interpretation of the claim, finding that "[c]laims directed to providing information based on feedback have previously been held patent-ineligible by this court" due to their being abstract. The Court further asserted that the "claims of the '412 patent implement a generic computer system to obtain functional results of providing information based on feedback from recipients" and that "Hyper Search does not show how the system of claim 1 is an improvement in computer capability." The Court further criticized the breadth and vagueness of the asserted claims, stating that this "sets the [them] apart from the claim language deemed sufficient to establish patent eligibility in the Federal Circuit's recent decisions, such as [Enfish v. Microsoft]."
Finding claim 1 abstract, the Court moved on to the second part of the Alice test. The Court noted that the claim "recites generic computer functionality such as a 'neural network module' and a 'server.'" Hyper Search's specification worked against it, as "[t]he specification states that neural networks were well-known in the art, and the inventors stated that the alleged invention is not limited to neural networks but rather to 'any artificial intelligence agent.'"
Hyper Search attempted to rely on BASCOM Glob. Internet Servs., Inc. v. AT&T Mobility LLC to establish that its claimed arrangement of elements provides a technical improvement. But the Court was not convinced since "Hyper Search does not identify any such example in the specifications in the patents-in-suit, and does not show how any of the prior elements are particularly arranged in the claimed inventions such that they improve the performance of the computer itself." Further, "Hyper Search does not identify any concrete improvements to specific computer functionality or solutions rooted in problems that only existed on computers, but instead generally asserts that each of the asserted claims is directed to solving a computer-related problem."
Consequently, the Court concluded that the asserted claims were invalid under § 101 and that the case should be dismissed.
Not unlike previous cases where the patent-eligibility of machine learning inventions was considered, the patentee here was hobbled by overly broad and non-specific claims that did not focus on exactly how the model operated or was trained. Additionally, there was no particular detail in the claim of how this model's input or output was processed. As a result, the facts of this case are so skewed against finding the claims eligible that it is a not a great data point. Thus, it would be a mistake to conclude that the outcome here does anything more than reiterate that generic claims with sweeping language are likely to have difficulties under § 101.
Ways to avoid or at least mitigate patent-eligibility challenges for machine learning inventions have been discussed previously and still apply. Until a claim written in accordance with these techniques is subject of a § 101 challenge, the eligibility landscape of machine learning inventions will remain largely unmapped.
