We are in the midst of an artificial intelligence boom. Widely considered the Fourth Industrial Revolution, AI is rapidly advancing computer technology from automated to autonomous.

Due to low-cost, high-capacity storage and computing power combined with the ubiquity of sensors capturing data of all types, companies are adding AI features to existing products and creating entirely new product offerings based in AI. "Big data" has created both the availability of robust training sets used to develop AI technology, and a need for technology that can process and filter large volumes of data for business applications.

The value and investments associated with the development of AI are trending upward. A recent McKinsey Global Institute paper estimates that AI can potentially create between $3.5 trillion and $5.8 trillion in value annually spanning nine business functions in 19 industries. Forbes calculates that there is already a 70 percent growth in business value in AI thus far into 2018.

AI technology has many embodiments, but artificial neural networks is one form of AI technology that companies are heavily investing in and implementing into their product offerings and businesses. This technology provides a vivid illustration of some of the challenges companies face when seeking to protect AI intellectual property. Modeled after the human brain, a neural network is an interconnected group of processing nodes that provides the architecture for a machine to learn how to perform certain tasks by analyzing training examples. For example, a deep learning, object recognition application may be fed thousands of identified images so that it can learn patterns associated with certain objects and learn how to identify them. Everyday applications including automatic translators and speech recognition applications are utilizing the deep learning techniques provided by neural network technology.

Recognizing the need to protect the value of the investment in AI, companies are increasingly securing IP protection for their development and application of neural networks. A company's legal strategy for protecting its AI IP should contemplate multiple forms of IP protection including patents, trade secrets and copyrights. Looking just at patent protection, the number of filings and patent grants in AI are on the rise. The U.S. Patent and Trademark Office, for example, is on pace to double the number of patents granted claiming "neural network" this year as compared to each of the previous five years.

While it is no surprise that IBM Corp., Google Inc. and Microsoft Corp. are the top assignees to be granted patents within the past five years with claims directed to neural networks, other companies outside of the computing industry are also developing new applications of neural networks and seeking patent protection. For example, Halliburton Energy Services Inc. owns several patents directed to drilling improvements that rely on neural network technology. Sports and fitness companies are also innovating with neural networks, with companies like Adidas AG and Fitbit Inc. receiving multiple patents in this area.

Protecting this investment in neural network IP requires navigating a complex and evolving legal landscape. One critical issue with respect to obtaining patent protection in the U.S. for applications employing neural networks is providing adequate support in the patent application to describe the invention.

When it comes to seeking patent protection for artificial neural networks inventions, satisfying disclosure requirements can present challenges. Underlying U.S. patent laws is a quid pro quo. In exchange for a limited monopoly via a grant to exclude others from practicing the claimed invention, an inventor must disclose to the pubic enough information about the invention to enable one of ordinary skill in the art to practice what is claimed.

Given the nature of some neural network inventions, meeting this requirement can be challenging. For example, when seeking protection for rule-based neural network systems, a research team may have developed rule sets that are effective for a specific application, such as speech recognition. Patent claims directed to a broader scope of applications may not be enabled by the rules developed. Accordingly, disclosing only those specific rules contemplated for one field of application may not satisfy the disclosure obligations of 35 U.S.C. Section 112.

The performance of AI embodied in an artificial neural network can depend on network topology (e.g., the number and types of layers, the number of neurons per layer, neuron properties, training algorithms, training data sets, etc.). The scope of the claims will depend on what the limited set of topologies disclosed in the patent teaches one skilled in the art to practice. In both the rule-based and network-based systems, where the systems have been developed heuristically, there may be questions regarding whether generalizations necessary to support the desired claim scope are disclosed in the patent. Millions of permutations of the network architecture or rules adaptable for various applications could exist. Disclosing a few and trying to define a broad claim scope may introduce risks. Providing a comprehensive disclosure laying out many embodiments may reduce some risk, but there are practical limitations regarding how many can and should be disclosed. This is an area where guidance may come from the pharmaceutical arts to understand the bounds of disclosure and written description compared to the number of permutations available balanced against the scope of the claims sought.

One strategy for addressing IP protection challenges for applications implementing neural networks may include functional claiming of invention building blocks, for example claiming elements as "means for classifying" or "means for responding to backpropagation learning." In addition to including in a patent application claim sets that recite specific structural details of a neural network invention, alternative claim sets that define claim boundaries by the functions that the elements perform could be beneficial. This type of functional claiming is authorized by 35 U.S.C. Section 112(f).

While functional claiming may have some strategic advantages, it does not entitle an inventor to claim elements functionally with the expectation of including all structure for performing the functions claimed. An inventor is only entitled to the structure that is disclosed for performing the claimed function and equivalents to what is disclosed. Also beware of attempting to satisfy disclosure requirements with "black box" schematics in a patent application with the expectation that one skilled in the art would be able to fill in the blanks. Recent court decisions have raised the bar requiring more details in some situations. Functions other than those commonly known in the art, often described by courts as requiring "special programming" for a general purpose computer, require disclosure of the algorithm for performing the claimed function. Functions known by those of ordinary skill in the art as being commonly performed by a general purpose computer or computer component, e.g., "means for storing data," do not require disclosure of additional supporting structure.

Accordingly, the ability to claim functionally does not obviate the need to disclose embodiments in the AI context.

AI penetration into businesses is already significant and growing rapidly both as a tool and as a focus of research and development for product offerings. The complexities of AI will require thoughtful strategies now to guide companies as they deploy and protect their AI-based innovations.

Previously published in Law360.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.