According to a new study just released by the Office of the Chief Economist at the USPTO (“Inventing AI: Tracing the diffusion of artificial intelligence with U.S. patents”), the number of annual AI applications filed at the USPTO more than doubled between 2002 to 2018, with broad growth across all technologies, inventor-patentees, and organizations. The USPTO performed the study to determine “whether AI technologies are growing in volume and, importantly, whether they are diffusing across a broad spectrum of technical areas, inventors, companies, and geographies.” See page 2.
Other key findings from the report:
- in 2002, the USPTO received around 30,000 AI applications, which were spread among 9% of the technology subclasses used by the USPTO,
- in 2018, the USPTO received over 60,000 AI applications, which were spread among 42% of the technology subclasses used by the USPTO,
- from 2002 to 2018, the percentage of filed patents related to AI grew from 9% to 16%, and
- since 2012, the largest growth in AI patents is in the areas of machine learning and computer vision.
The key challenge faced by the UPSTO in generating their findings was to determine whether a particular patent document is directed to AI. Interestingly, the USPTO used artificial intelligence to make this determination. In particular, the AI-patent data provided in the study was extracted from a database of approximately 8.5 million patent documents (patents and patent applications published between 1976 and 2018) using a machine learning method in which each of the patent documents was evaluated by eight separate machine learning classifiers (neural networks) to determine whether the document is directed to one or more of eight different AI component technologies.
Specifically, the USPTO defined the eight different AI component technologies as: knowledge processing, speech, AI hardware, evolutionary computation, natural language processing, machine learning, vision, and planning/control, and then trained eight neural network classifiers, one for each AI component technology, using training data prepared for each AI component technology.
After evaluating all 8.5 million patent documents against each classifier, the USPTO used metadata associated with each patent document, such as the filing date, CPC classification, inventor-patentee, etc. to extract the study data for the graphs illustrating the growth and diffusion of the AI component technologies over the past forty years.
For example, to derive the “volume” data of Figures 2 and 3 in the report, the filing date associated with each AI-classified patent document was used to determine the number of AI patents filed in a given year. Note that a patent document could be classified as AI by more than one of the different AI component classifiers, and was counted as an AI application if it was classified as AI by at least one of the different AI component classifiers.
As another example, to derive the “diffusion” data for Figure 4 in the report, the CPC classification data associated with each AI-classified patent document was used to determine which USPTO technology subclasses had at least one AI patent in a given year.
Each of the eight AI component classifiers used by the USPTO is a neural network having 664 input nodes, a hidden layer of 64 neurons, and a single output node. The numerical value of the output node indicates whether the input data representing a particular patent document belongs to the corresponding AI component or not.
The 664 inputs to each of the eight component classifiers is derived from portions of each patent document, including the text of the abstract (300 inputs), the text of the claims (300 inputs), and patent citations in the document (64 inputs). For example, the text of the abstract and claims was first preprocessed to be all lowercase, and to remove starting numbers, symbols, formulas, and extra spaces, and then concatenated into a text string. Further, each of the claims and abstract text strings, after preprocessing and conversion to a vector, is fed into a separate long short-term memory (LSTM) neural network with a 300-dimensional output. The citation information was similarly preprocessed and fed into a separate network with a 64-dimensional output. After this initial processing, the 664 inputs were fed into the primary neural network of the component classifier to generate the single output value.
Each component classifier was trained with a set of seed patents known to belong to the particular AI component technology, as well as a set of anti-seed patents known to not belong to that component technology. For example, for the machine learning component technology, the USPTO first identified 959 seed patent documents, and then looked at several generations of backward and forward citations from those seed documents (and family members) to determine roughly 470,000 other documents that were potentially related to the seed documents. Finally, 15,000 anti-seed documents were selected such that none of the anti-seed documents were selected from the 470,000 related patent documents.
After each of the component classifiers was sufficient trained on the seed and anti-seed patent documents, and the classification results validated, each of the 8.5 million patent documents was evaluated by each of the eight component classifiers. As noted above, a given patent document could be classified as belonging to more than one of the AI component technologies.
An overview of the machine learning methodology described above is provided in a short Appendix to the report. In addition, a 53-page supplementary document provides many more additional details. See https://www.uspto.gov/sites/default/files/documents/OCE-ai-supplementary-materials.pdf.