Patent applications have to navigate a minefield of potentially fatal pitfalls, especially when the invention is related to systems utilizeing smart algorithms or related software to solve thorny technical problems. While progress has been made at the USPTO, there are still many examiners and supervisors utilizing every tool available to reject applications. And even if you survive the patent office, those same issues will likely be prevalent again in an IPR or in district court litigation.
One particular issue unique to artificial intelligence cases, including those utilizing deep learning such as neural networks, is that of training. An initial strategic decision is whether the training aspect is even something that should be claimed, as often this occurs on the back end and can be difficult to observe or identify from an infringement point of view. Of course, there are cases where it may be important to actually claim aspects of the unique training approaches.
A recent PTAB case illustrates the difficulty that can arise in claiming training aspects for AI tools with respect to written description. Some examiners (and PTAB judges) can be exacting in their requirements for adequate written description in terms of details explaining exactly how something is accomplished. The claims in an example case were related to a wearable article for determining a type of physical activity in which an individual wearing the wearable article has engaged and for initiating direction of an advertisement to the individual tailored to the determined type of physical activity. The invention utilized a neural network trained based on accelerometer data as follows:
wherein the controller includes a pattern recognition algorithm implemented via a neural network configured to determine a type of physical activity, from among a predefined set of physical activity types, through analysis of the output of the three-dimensional accelerometer, the neural network is a trained neural network trained based on acceleration data obtained through performance of each of the predefined set of physical activity types;
* * *
wherein the controller is further configured to … use the trained neural network to analyze the measurements of acceleration over time for the extremity of the individual along three axes and to determine the type of physical activity in which the individual has engaged, from among the predefined set of physical activity types,
The examiner asserted the claims lacked sufficient written description as follows:
With regard to the determination of activity type and "activity signature patterns" -while the disclosure generally suggests pattern recognition and activity type classification using a neural network trained with data obtained from performing a plurality for activities (, -, [01531), this level of disclosure amounts to a general approach to programming the computer and a suggestion of possible inputs ( e.g. speed, displacement, shock, pulse, temperature, etc.) and desired result (e.g., an accurate activity type classification based on new input). However, there are several additional algorithms or steps that would be required for a practitioner (and Applicant) to perform to actually program a computer to perform these functions. These additional algorithms or steps would have a significant impact on the actual program (and quality/accuracy of results) a practitioner would end-up with. While Applicant is not required to actually submit the source-code of their programming, sufficient disclosure of the algorithm/steps requires a sufficient disclosure of all algorithms/steps that have a significant impact on Applicant's actual program/model (and the actual program a practitioner would end-up with if attempting to recreate Applicant's method/system).
Here, the applicant pointed to general disclosure in the specification as to how the acceleration data can be integrated to obtain velocity and position, and differentiated to obtain shock. Further, the applicant explained how various combinations of these can indicate jogging vs. walking, etc. However, the specification did not have any detailed examples of specific algorithms, including specific training algorithms. The PTAB found this to be insufficient since the specification only described components of a pattern recognition algorithm, not a complete algorithm:
The adequacy of written description is a question of fact and is not shown by mere argument (citations omitted). On this record, we agree with the Examiner that the Specification fails to provide adequate written description of … how the neural network is trained based on the obtained acceleration data. Absent evidence of what a person of ordinary skill in the art would have known or been able to determine from the Specification, we are not persuaded that the Specification provides adequate written description to support the claims.
So, we see here yet another example of where short-cutting the detailed description, even on what might seem like tangential aspects not important to the core invention, can lead to written description problems that are fatal to the application as claimed.