In the dimly lit corridors of the European Patent Office, where paper trails stretched like the winding paths of an ancient labyrinth, a solitary application whispered its technical aspirations into the ears of examiners. The application, filed by Mitsubishi Electric Corporation, bore the solemn title: “Hierarchical Neural Network Device, Learning Method for Determination Device, and Determination Method.”
The problem it sought to solve was an affliction known to many in the realm of artificial intelligence: the curse of computational excess. Conventional neural networks, bloated with an inordinate number of connections, suffered under the weight of their own complexity. They required massive computational resources and were prone to overfitting, becoming nothing more than mirror-walled chambers that echoed the nuances of their training data rather than generalizing knowledge to unseen cases.
Mitsubishi’s solution was one of elegant sparsity. Instead of forging a fully connected network, the invention proposed a hierarchical structure governed by a sparse parity-check matrix, akin to those used in error-correcting codes. By imposing predetermined, fixed connectivity constraints, the neural network would achieve high-speed learning while maintaining robust discrimination performance. This departure from the standard practice of data-driven connectivity formation was radical, a statement that perhaps the very foundation of learning could be shaped by mathematical elegance rather than empirical messiness.
Yet, the Examining Division of the European Patent Office was unmoved. In its judgment, the invention’s distinguishing feature—its sparse connectivity matrix—was but a mathematical abstraction, an ethereal concept drifting in the non-technical ether. It did not serve a specific technical purpose, nor did it bring about a tangible modification to the machinery that executed it. Instead, it remained trapped within the ivory tower of mathematical methods, condemned to exclusion under Article 52(2) EPC.
Undeterred, Mitsubishi appealed. The appeal was a desperate entreaty to logic, armed with precedents and a resolute belief in the technical nature of their invention. They argued that their hierarchical neural network served a technical purpose, akin to how encryption, an almost purely mathematical endeavor, was recognized as technical because it enhanced security. They leaned on jurisprudence, invoking decisions like T 1326/06, which acknowledged that methods for encoding and decoding data could indeed serve a technical function. Furthermore, they pointed to real-world ramifications: their structure reduced storage and computational costs, a tangible and practical advantage.
The Board of Appeal, seated in its high court of intellectual scrutiny, examined these arguments with a dispassionate gaze. It acknowledged that neural networks, at their core, were mathematical entities—structures that defined classes of functions, each instance differing only in its learned parameters. To rise above the mire of non-technicality, a neural network had to solve a concrete technical problem, not merely refine an abstract function.
Mitsubishi’s claim, the Board reasoned, did not specify a real-world application where the purported efficiency gains would manifest as a practical, industrial advantage. The reduction in computational resources, though significant, was an incomplete argument. One could not simply compare a sparsely connected neural network to a fully connected one and claim superiority; rather, the question had to be whether the trade-off—reduced complexity at the cost of potential loss in learning capacity—constituted an inventive technical contribution. Here, Mitsubishi had fallen silent.
Moreover, the Board rejected the analogy to cryptography. In cryptography, the technical problem was clear: securing digital communications against unauthorized access. In contrast, neural networks, particularly those dealing with classification and regression, did not inherently solve a technical problem unless the output had an implied technical use. Without a clearly defined technical application, the invention remained within the boundaries of mathematical abstraction.
Ultimately, the Board ruled that the claims did not satisfy the inventive step requirement under Article 56 EPC. The appeal was dismissed, and Mitsubishi’s neural network was left stranded in the liminal space between mathematical ingenuity and patentable invention.
For future applicants navigating these treacherous waters, the lesson was clear: a neural network, no matter how novel in structure, must be tied to a specific technical implementation or application. The computational advantages it offers must translate into a concrete technical benefit, be it in hardware efficiency, processing improvements, or a well-defined industrial process. Otherwise, the invention, like so many before it, would vanish into the void, another echo in the silent halls of the European Patent Office.
Based on T 0702/20 (Sparsely connected neural network/MITSUBISHI).
Key Takeaways from the Decision
- Defining a neural network solely through its mathematical structure, like using a sparse parity-check matrix for connections, is considered a mathematical method and is excluded from patentability as such under European patent law.
- While implementing a neural network on a computer gives it a technical character, this isn’t enough for patentability; the specific features must solve a technical problem.
- Claiming reduced computational resources or storage due to a different network structure is not automatically a technical effect if the change fundamentally alters the network’s ability to learn and perform without a demonstrated technical advantage.
- General claims about neural networks automating tasks are insufficient; the patent application needs to specify the particular technical problem being solved and the type of data involved to show a technical purpose.
- Claims of technical benefits like improved learning or preventing overfitting need justification and evidence, especially when the network structure is fixed before training and independent of the data.