Neural networks and sets of weights are further technical innovations that fail to fit established categories of intellectual property law. Neural network weights fail to neatly match the requirements of copyright law (intended to protect text against copying) or patent law (intended to protect machines against unlicensed manufacture, use, or sale). Nevertheless, in this brief paper, intended for engineers and managers investigating intellectual property protection for neural nets, Stern shows that neural networks and weights can be protected by US patent law, but weights probably cannot be copyrighted.

Stern succinctly describes the basic requirements of both legal schemes and then applies them to neural networks. On the patent front, Stern gives a practical example of how artfully worded patent claims should be able to protect a neural network program. The claims must follow the customary charade (necessary to convince the US Patent & Trademark Office that the claims qualify for protection) of reciting a pure software system as an “apparatus” comprising various hardware elements such as a “processor,” an “input device,” and so on. Unfortunately, Stern’s paper appeared just a month too soon to discuss the recent landmark decision of the Court of Appeals for the Federal Circuit, In re Alappat [1]. In Alappat, a patent examiner rejected a patent application of Textronix for a method of displaying an oscilloscope trace using antialiasing to smooth the waveform, claimed as a “rasterizer apparatus” comprising four “means” for performing certain functions. On appeal, the Federal Circuit reversed the examiner and reinstated the application, holding that the claims recited a legally cognizable “machine.” Since the US Supreme Court rarely hears patent cases, the decision is expected to stand.

Although the Federal Circuit issued a split decision (of the 11 judges, 6 voted to approve the claims, 2 voted to reject them, and 3 abstained), many in the patent bar believe that Alappat is a watershed decision, declaring once and for all that software inventions recited as “machines” are patentable. The decision supports Stern’s hypothesis that a novel and nonobvious neural network program is patentable; its omission from Stern’s paper, however, underscores that the interested reader cannot rely on dated material in this area.

Turning to copyright, Stern focuses on the problem of protecting weights, since there is no question about the protectability of a program to implement a neural network. For a set of weights, he finds no guarantee that copyright will provide protection, for two reasons. First, a set of weights does not meet the statutory definition of a copyrightable program, which must be a set of functional statements or instructions. Second, a set of weights does not qualify as a human “work of authorship”; others have argued that “computer authorship” should be protected, if only to prevent inconsistent application of the copyright law [2].

A more important question is treated in just three paragraphs: whether protecting the weights of a neural network adequately protects the value of a trained neural network. Stern’s brevity on this issue is disappointing, both because he is addressing a technical audience and because so much legal scholarship on the protection of software products ignores the critical question of what management wants to protect. Anyone venturing into this field must answer that question long before turning to Stern’s paper for an overview of the legal issues.