Council Post: How To Patent Artificial Intelligence And Machine Learning Models

So, you’d like to patent your artificial intelligence (AI). Congratulations, it’s going be a rough one. Broadly speaking, patents can be afforded to systems, apparatuses and processes. So why is patenting AI rough? Because patent language describes inventions in terms of what does what, why, when and how. In other words, you need to describe your AI invention in terms of structure. But U.S. courts do not consider computer structure (e.g., processors, computer memory and the like) to be machine structure, believing computer structure to implement only functions rather than execute any transformations of product. As a result, many machine learning models seeking patents in the U.S. are endangered.

That doesn’t mean you should give up on patenting your AI. Rather, it’s important to understand how your AI will be approached from a patenting perspective. Here are some tips for assessing the major categories of machine learning systems from a patenting perspective and the types of information you should give your patent counsel.

You want to patent a supervised learning model.

Here your system learns from a teacher that is a human. The most common example of this type of AI is a training model having a classification task, such as an image recognition model that trains the AI to predict what is in an image. Here, the teacher needs to provide the system with many examples of the subject, including both predictors and labels.

Now think in terms of structure, as structure is what you should claim in your patent. Without structure, your patent application relies mostly on functions, which many United States Patent and Trademark Office (USPTO) examiners will disregard as lacking patentable weight. So, what are the structures in a supervised learning model such as an image recognition model?

• Input structure you should provide to your patent counsel:

Does your system have any machine inputs, such as scanners, sensors or cameras that are detecting images on an ongoing basis? How often do the inputs assess their subjects (is it automatic or machine controlled?)? What is their range of detection? Is detection triggered by an event? The downside of this type of system is that the input is most likely a human, but maybe the human works with a machine input, which can help your patentability case.

• Transceiver/receiver structure you should provide your patent counsel:

After data detection, where does the data go? Is it uploaded to a cloud computer or downloaded to a local network? Does uploading/downloading occur automatically, or is it controlled by a computer, and why?

• Storage structure you should provide your patent counsel:

Most attorneys have form language for this part and will throw all the computer memory and processor terms into your patent application. But I have had clients inform me that storage of data being on a cloud server versus a local server affects the workability or efficiency of an AI model. This is the type of information you should provide, as well.

• Processing structure you should provide your patent counsel:

This is the tricky part. How is the outcome or output (e.g.