Obtaining strong patent protection for AI inventions is an increasingly more difficult task. Not only is the volume of AI patent filings (and thus the corpus of prior art) growing exponentially, but AI inventions face additional hurdles in terms of Section 112 (written description/possession) and Section 101 (abstract idea). If all of that was not enough, the USPTO has further tricks up its sleeve in terms of claim interpretation - particularly when it comes to claims related to AI training innovations - that enable easier rejections.

Before elaborating on the above issues, some background is provided as to product-by-process claims. Understanding the nuances of these claims is important in both drafting and prosecuting patents, and as shown below, particularly relevant for AI innovations.

A product-by-process claim is a type of patent claim that defines a product (e.g., a physical object or composition) in terms of the process used to make it. Unlike traditional product claims, which define the product by its physical characteristics or structural features, product-by-process claims describe the product in relation to the steps of the process employed in its creation. For example, instead of claiming a "device" based on its structural features or chemical composition, a product-by-process claim might describe a "device" as one that is produced by a specific sequence of manufacturing steps.

According to MPEP 2113, the examination of product-by-process claims generally follows the principle that the process steps do not themselves constitute part of the claim's definition of the product. Instead, the focus is on whether the claimed product, regardless of the process of making, is new and non-obvious over the prior art. The key consideration here is that the product itself must be distinct and not merely the result of a known process. MPEP 2113 confirms that even though product-by-process claims are limited by and defined by the process, determination of patentability (as opposed to infringement) is based on the product itself.

Patents for software innovations introduce an interesting twist to product-by-process claims. Although traditionally associated with physical products, the concept of product-by-process has been extended to software-related inventions under certain circumstances. For example, if a software product (such as an algorithm or executable code) is defined in the claims by the process used to generate or compile it, it might be examined through a product-by-process lens. Likewise, a device with AI software trained by a specific process can invoke a product-by-process interpretation as to the special training (thus enabling an examiner to ignore such features).

Patent practitioners need to be aware of this recent trend by not only some examiners, but also by the PTAB. A recent case from NVIDIA illustrates the potential trap for applicants, as the PTAB can raise the issue, sua sponte, in an appeal decision. See Appeal 2023-001414, Application 16/752,225.

Claim 1 on appeal is reproduced below:

1. A processor, comprising: one or more arithmetic logic units (ALUs) to determine a 3-D pose from an image using one or more neural networks, the one or more neural networks trained by at least: 

obtaining a 2-D image of an appendage; generating a proposed 3-D pose of the appendage from the 2-D image of the appendage;

determining one or more losses that are based at least in part on a model that describes allowable appendage positions; and 

adjusting the one or more neural networks based at least in part on the one or more losses.

Here, the invention relates to a particularly trained neural net. During prosecution, the examiner relied on a rejection that was lacking as to sufficient disclosure in this regard. On appeal, the PTAB dismissed this issue by interpreting the claims as product-by-process as to the training. From a brief review, it appears that the examiner had not previously raised the product-by-process issue.

From the PTAB decision (citations omitted):

[Claim 1] sets for that the neural networks used by the ALUs have previously been trained by the recited steps. That is, claim 1 is not directed to the steps for training the neural network. But rather, the training steps constitute product-by-process language. The law pertaining to the product-by-process doctrine is clear:
[E]ven though product-by-process claims are limited by and defined by the process, determination of patentability is based on the product itself. The patentability of a product does not depend on its method of production. If the product in a product-by- process claim is the same as or obvious from a product of the prior art, the claim is unpatentable even though the prior product was made by a different process.
Appellant does not dispute whether Onen teaches or suggests ALUs that determine a 3-D pose from an image using one or more neural networks. Appellant only argues that the Examiner has not established that the trained neural networks of Onen are trained by the process recited in claim 1. But Appellant has not provided persuasive evidence that training a neural network by the method set forth by the product-by-process language necessarily will produce trained neural networks that are tangibly distinguishable from neural networks that are trained by other processes. Accordingly, Appellant has not demonstrated reversible error in the Examiner’s obviousness rejection of claim 1.
We therefore affirm the obviousness rejection of claim 1. …

Product-by-process claims represent a nuanced area of patent law that can impact both traditional and AI-related inventions. By understanding the general rule for examination and the implications for AI inventions, practitioners can better navigate the challenges of drafting and prosecuting such claims. As always, careful consideration and strategic planning are essential. As is often the case with appeals, patent practitioners will do well to anticipate potential new interpretations by the PTAB that can enable an easy way to affirm.