This simple plot reflects a profound truth:

This plot involves the field of speech recognition, which software developers have pursued since the 1970s through the traditional approach of algorithm design. The product of 50 years of research and experimentation in audio processing, phonemes, and language models produced algorithms with a word error rate around 20%, with moderate but slow improvement based on extensive training.

Over the past several years, data scientists have adopted a new approach of submitting certain aspects of speech recognition to machine learning models, such as deep neural networks (DNNs). The results are surprising: as compared with developed algorithms, machine learning models of speech recognition exhibit both an initially lower error rate and faster improvement from additional training. A recent article published by the Association of Computing Machinery notes: “As a result of the success in deep learning, speech recognition researchers are returning to using more basic speech features such as spectrograms and filterbanks for deep learning, allowing the power of machine learning to automatically discover more useful representations from the DNN itself.”

This example is one of many scenarios in which machine learning techniques are demonstrating proficiency in complex tasks that exceeds not just the capabilities of human experts, but of computers that are programmed with human-designed algorithms. Machine learning is gaining acceptance as a powerful set of tools that are broadly applicable to the field of automation – one that complements and, in some cases, surpasses algorithmic solutions.

However, the maturation of machine learning coincides with a tumultuous phase of U.S. patent law involving patent-eligible subject matter. Countless articles have documented the tension among the federal courts, USPTO administration, and even Congress as to the delineation between patent-eligible improvements in computing technology and concepts that are deemed “abstract” or “conventional.” Indeed, the characteristics of machine learning that yield such technically compelling power may conflict with the understanding of patent examiners and judges as to the metes and bounds of patent-eligible technology.

Nevertheless, the demand for patent protection of machine learning inventions is escalating at an unprecedented rate. For example:


(Click here to see the animated graph)


Furthermore, this demand is not limited to the art units and CPC classes that are focused on machine learning, or even computation in general. On the contrary, applications that discuss machine learning are proliferating across a wide range of technology areas:


(Click here to see the animated graph)


Some examples of CPC classes in which machine learning unexpectedly appears in patent applications include:

  • G08G: Traffic control systems (678 applications projected for 2019)
  • G09B: Educational appliances; communication appliances; planetaria; globes; maps (387 applications)
  • H04R: Loudspeakers; microphones; gramophones pickups; public address systems (220 applications)
  • E21B: Earth drilling (186 applications)
  • H01Q: Radio antennas (162 applications)

The computational power of machine learning and the obstacles to patenting such inventions provide fresh challenges for both technology companies and patent practitioners. These circumstances require us to pay close attention to the rapidly evolving technology of machine learning and its industrial deployment; the growing demand and strategies for the protection and commercialization of machine learning techniques; and the fluctuating state of patent law that may affect the options for protecting machine learning inventions.