Can we patent our machine-learning innovation?

Machine learning refers to a broad range of techniques with which a computer makes determinations about input data without explicit programming on how to get the result. Rather, the computer is programmed to "learn" to make determinations from features of the data. These techniques can be incredibly powerful and have been used in a wide range of industries, from guiding investment by predicting financial market behavior to improving the accuracy of voice interface systems. The business value of a new machine learning solution may be clear, but can it be patented?

The phrase "on the internet" became a running joke with patent attorneys after the dot com boom. That period saw numerous applications filed (and patents issued) for providing a service using a computer connected to the internet. The primary arguments presented for novelty often focused on generic use of the computer and internet rather than the details of the service itself. In fact, in many cases, the computer provided the service in essentially the same manner as previous manual approaches.

Arguably, this was a driving factor behind the Supreme Court's ruling in , 573 U.S. 208 (2014), which invalidated patent claims as merely directed to an abstract idea. The court had many years previously held that abstract ideas are not entitled to patent protection under 35 U.S.C. Section 101. Over the last five years, courts have used the Alice precedent to invalidate numerous computer- and internet-based patents. Although many have argued the Supreme Court should have instead invalidated the claims as obvious under 35 U.S.C. Section 103, the approach adopted does hold a certain appeal for courts. Determining that a claim is not directed to patentable subject matter under Section 101 is primarily a matter of law. Thus, courts can avoid potentially complex factual analyses of whether performing known processes by computer (or "on the internet") would have been obvious to one of skill in the art at the time the patent application was filed.

Since the Alice decision, patent attorneys have developed strategies to convince the Patent Office and the courts that their claims are not directed to abstract ideas. Generally, claims that focus on specific details regarding how to achieve the desired outcome using a computer and the internet are far more likely to be patentable. In particular, patent attorneys draft claims that highlight differences between computer-based solutions and manual approaches and explain how using the computer provides tangible benefits over existing approaches. These practices are reflected in the latest examination guidelines, issued by the Patent Office in January 2019, which state a claim is not directed to an abstract idea if it is integrated into a practical application.

In many ways, machine-learning inventions are the modern-day "on the internet." Inventors often initially use the phrase "using machine learning" in describing the novelty of their inventions. But the use of machine learning is now well-known. A claim to obtaining a result using a computer without something more will almost certainly be rejected by the Patent Office. Whether the claim is rejected as an abstract idea under Section 101 or as obvious under Section 103, merely adding that the result is achieved using machine learning is unlikely to be enough. That is not to say that machine-learning innovations are not patentable. Rather, when faced with such an invention, one must dig deeper to determine how the machine-learning is implemented, and why that represents an advancement over the state of the art. The following areas are good candidates for identifying an inventive concept to support patentability: training data generation, data preprocessing, and model structure.

For supervised learning solutions, having good training data is key. The training data must be labelled correctly and needs to provide reasonably broad coverage of the possible range of inputs. In many cases, this task is non-trivial. For example, consider using a set of photos uploaded to a photo-sharing website to train a model to identify those that include dogs. Photos tagged by the uploader as including dogs are likely to include dogs, but one cannot know if other photos include dogs without inspecting each photo individually. If the number of photos is large, this is impractical. Engineers often use innovative techniques to generate reliable training data, and such techniques may be entitled to patent protection.

Preprocessing is another rich area for identifying patentable features. The effectiveness of a machine-learning solution is often dependent on the form of the input data. Because machine-learning classifies based on correlation, without any underlying knowledge of causation, there is no guarantee that the output has real-world meaning. Therefore, preprocessing the input to remove potentially misleading information often improves results. For example, a classifier trained with the intent of distinguishing between photos of checkers or chess may instead learn a correlation in the training set between background color and which game is depicted. This correlation is likely nothing more than random chance and is specific to the training data rather than representing a more general rule. By converting the images to black and white, this error may be avoided, forcing the classifier to focus on more relevant features of the images. Novel preprocessing that improves the effectiveness of the model may be entitled to patent protection.

In other instances, the structure of the model itself may be patentable. In many cases, machine-learning is implemented using an off-the-shelf model. However, in certain instances, an engineer may modify the structure of the model itself to build a functional system. For example, neural networks contain multiple layers of nodes, with different layers performing different functions. If the particular number and arrangement of layers provides a significant contribution to the functionality or efficiency of the system, the structure of the model itself may be patentable.

In conclusion, merely applying machine-learning to a data set is unlikely to be entitled to patent protection, but specific training data generation methods, preprocessing techniques, and model structures that are tailored to solving a particular problem often are. If it took significant effort to build a machine-learning solution that works as intended, consider consulting a patent attorney to explore your options.