Virtual assistants like Alexa and Siri can answer questions, manage calendars, and translate languages. Self-driving cars accelerate, brake, and steer, and someday soon will operate without human intervention. Smart medical devices track vital signs and notify doctors of oncoming medical problems. Banks detect unusual purchases and automatically activate anti-fraud measures. Each of these previously impossible ideas are now a reality thanks to artificial intelligence and machine learning – and the University of Tennessee is at the forefront of advancing these exciting new technologies to tackle some of today’s most challenging problems.

Artificial intelligence (AI) is essentially the ability of advanced computer systems to think and make decisions in a manner similar to humans, perceiving and understanding their local environment. For example, self-driving vehicles and autonomous robots not only observe everything around them, but also predict the future actions of those same objects. As a driver, we all know that a single blinking light on the back of a car means “turn signal” and indicates a (likely) change of direction. We also know that two blinking lights mean “hazard” and indicate completely different expected behaviors from that vehicle – it may slow down, stop, swerve, or pull off to the shoulder of the road. An effective AI driving system must make these same assessments.

Machine learning (ML) is a subset of artificial intelligence; it refers to computer learning through pattern recognition rather than explicit instructions.

In a typical ML operation, the computer analyzes supplied training data in order to build a mathematical model that can make predictions and decisions. Returning to our self-driving car example, this training data might include video and sensor readings collected from cars and trucks driven millions of miles.

UT professors and researchers across the state are taking the potential of AI and machine learning and applying these techniques to address challenging problems in medicine, engineering, and computer science. At the Health Science Center in Memphis, Director of the Center for Biomedical Informatics Bob Davis  and his team present ways to improve the detection and prediction of sepsis and cardiac arrhythmia. Meanwhile, UT’s TENNLab team in Knoxville develops neural network architectures and dynamic training algorithms for high-performance, low-energy computing that emulates the human brain, known as neuromorphic computing. Many more applications of AI and ML appear throughout recent UT publications, theses, and dissertations.

As more technologies incorporate AI and machine learning, universities and businesses around the world grapple with a central question: how to protect this intellectual property.

In the past year, patent offices in the U.S., EU, and Japan issued guidance clarifying the scope and bounds of patentability for inventions based on AI and machine learning. In November, the European Patent Office (EPO) published updated Guidelines for Examination, which define the key patentability question: “Is the claimed invention directed to a mathematical method, or is it tied to a control or operation system and has a technical character?” Under these guidelines, the use of a neural network (a type of mathematical algorithm) in a heart-monitoring apparatus is considered a patent-eligible invention, but the use of a similar algorithm to simply classify data records without a specific technical use would be ineligible.

The guidance from the Japanese Patent Office (JPO) teaches applicants to match the scope of the patent claims against the written description of the AI/ML invention. Because these techniques are powerful and flexible, broad claims based on foreseen applications are easily made. However, if the description in the patent application is limited to one specific example, the scope of the claims must follow the same limitations. The example provided by the JPO is an AI-based system for predicting whether an unknown chemical will induce a skin allergy. The system was trained on skin allergy data from thousands of known compounds, but the patent application claimed a system both for predicting skin allergies and for predicting any other allergic response. As the written description’s example specified training the system only for skin allergies, the scope of the broader claims about all allergic responses exceeds the enabled description, making it ineligible.

Like the EPO guidance, the U.S. Patent Office (USPTO) uses examples from several technical fields, such as facial-detection methods and network-monitoring algorithms, to demonstrate how the office will analyze subject-matter eligibility of AI/ML inventions. The USPTO guidance teaches that a claim is patent eligible if it:

  1. Is more than just a mathematical correlation, mental process, or business method, and;
  2. Is integrated into a practical application, or uses non-routine and non-conventional steps.

With a stronger understanding of the patentability of inventions based on machine learning and/or artificial intelligence, UTRF is now well-positioned to support UT researchers who develop innovative applications of these technologies. Have an idea or invention that you think may have commercial potential? Contact us today at utrf@tennessee.edu or idea.tennessee.edu.