Like many insights-driven organizations, the United States Patent and Trademark Office (USPTO) leverages data analytics and technologies such as AI and machine learning (ML) to increase the efficiency and performance of its operations and to improve the quality of systems and processes.
While AI and ML algorithms are critical to the agency’s endeavors, the government agency’s guiding principle is to take a human-first approach in developing and using these technologies to refine and scale its initiatives. AI and ML tools help empower the work of human experts and augment their ingenuity in the work they do, but at this point they can’t match the subtle nuances or reasoning capabilities of the human mind, notes USPTO CIO Jamie Holcombe.
To supplement the technology, the agency relies on input from thousands of experienced workers, captured passively and actively, to train and refine AI-driven models to ensure the technology delivers expected outcomes. The agency has awarded over 11 million patents since its founding and employs more than 12,000 people, including engineers, attorneys, analysts, and computer specialists. A continuous flow of feedback from its patent examiners on the front lines is also used to improve AI/ML models to fuel the development of new products and support activities in two key areas: patent search and classification.
Doing a comprehensive patent search can be challenging given the explosion in the volume of data and possible sources of “prior art,” notes Holcombe. To meet the challenge, technology teams are rolling out an AI component in a new patent search tool to help examiners find the most relevant sources they need as they scrutinize applications. This is important because each one of the more than 600,000 applications received yearly by the USPTO on average contains approximately 20 pages of text and figures, or roughly 10,000 words describing the claimed innovations. The agency’s IT organization also developed and deployed a classification tool that identifies and matches the classification symbols associated with an invention from over 250,000 possible categories.
In both cases, the models were developed and are continually enhanced by input from human experts who provide a human touch to determine whether something is truly new or novel, and then apply law, facts, and expertise to reach a decision.
Exploring human channels in the information stream
Having a constant flow of feedback from examiner experts and others may be an asset, but it is not the only route the USPTO is taking to identify new channels for innovation and global expertise to help solve important challenges and scale AI. Earlier this year, the agency turned to the AI research community and Google Kaggle, a preeminent technical and social platform used by data scientists and others to exchange thoughts and ideas. It launched a worldwide global coding competition in March, offering $25K in prize money and calling upon AI researchers and data scientists to write code for evaluating the semantic similarity of phrases.
The competition attracted more than 42,900 entries before closing on June 30 and involved 1,800-plus global teams working together and leveraging publicly available patent data sources. The goal of the competition was to move the needle on understanding patent language with AI for the agency and for the patent community, explains Holcombe. “The result will not only be better phrase algorithms for patent search, but the winning models will become part of the public domain,” he says.
The USPTO also made use of other public information resources such as Golden, a free ‘Wiki-style’ AI/ML-driven platform launched in 2019 that scours the Web to match topics with relevant and available data, pulling it together into a stream of information. An AI algorithm, working behind the scenes, continues to add related data whenever it becomes available. Anyone can seek information on companies, their patents, and funding sources such as venture capital.
The A, B, C’s of an AI/human alliance
While volumes are written about technology convergence, taking a ‘human-centric’ approach to AI and ML development can be challenging given the diverse and complex distinctions of human nature. To keep efforts on track, the USPTO, under Holcombe’s direction, developed a guide for progressing from pilots to prototypes to production. The alphabetical basics of that guide are the following:
A is for alignment: There must be a strong nexus between the business and IT staff, says the USPTO’s IT chief. “The best cross-functional teams are composed of technical staff working side by side with business representatives, all within an agile environment that promoted planning, doing, checking, and adjusting.” Agile, and/or “DevSecOps” practices rely on swift moves, transparency, and a product mindset. To maximize progress, leaders engage early and often with their teams and stakeholders.
B is for business value: Start with a business case that has obvious value for a core, strategic operation. Such a use case should address a challenge where AI and ML can logically help. “As a 100% fee-funded agency, our teams approach technical challenges though a rigorous business and ROI lens,” Holcombe points out.
C is for customers (and employees): AI/ML solutions are designed to augment examiners and other subject matter experts, rather than to replace them. So, emerging tech teams test and adjust concepts with internal customers before, during, and after any launch. Examiners who use the products help drive AI innovation, with some of them “on detail” and working side-by-side in the Office of the CIO to provide critical input. “Because we weave our customers into the process early, we get powerful feedback which helps drive adoption,” Holcombe notes. “Also, customers keep us honest in deploying AI that is accountable to agency experts and the public we serve.”