As a foremost component of artificial intelligence (AI), machine learning is based on computational models that learn their own parameters. Such models include decision trees (e.g., random forest), linear regression, nearest-neighbour or Gaussian process predictors, as well as artificial neural networks (ANNs).
ANNs are computational models inspired by biological neural networks. They have successfully been applied to, e.g., speech recognition, text processing, and computer vision. Several types of neural networks have been investigated, such as feedforward neural networks (e.g., convolutional neural networks), recurrent neural networks, and transformers for large language models (LLMs), to cite a few.
As research intensifies, ever larger models are being deployed, starting with so-called foundation models, which are trained on broad data and are used across a range of use cases (text, images, music, videos, astronomy, radiology, robotics, genomics, coding, and mathematics).
ANNs may be implemented in hardware, e.g., as crosspoint devices (crossbar array structures of in-memory compute accelerators) or reservoir networks. However, ANNs are mostly implemented in software.
A recurring question is whether AI innovations are at all patentable. Computer programs are not patentable as such. However, can practical applications or implementations of AI be patented? The answer is yes, provided that the novel features of such inventions have technical character. This is a touchy point but, fortunately, European jurisdictions including Switzerland have a relatively clear legal framework for such inventions.
The current situation of the patentability of inventions relating to AI and machine learning in Europe can be summarized as follows.
- Hardware innovations (e.g., neuromorphic processors and in-memory compute accelerators) are patentable, like any other machine. What poses a problem are inventions based on computational models (e.g., neural networks) executed by conventional processors.
- Still, as with mathematical methods in general, AI-related inventions are potentially patentable if they involve a specific technical application (e.g., the identification of irregular heartbeats) and/or a specific technical implementation (e.g., exploiting parallelism or vector processing).
- There remains the examination of inventive step, for which it is necessary to (fiercely) defend the indirect technical contributions of abstract features (e.g., algorithms). There, interactions between abstract and tangible features will be decisive.
Obtaining patents for new implementations or applications requires great care. It is important to carefully consider all technical aspects (e.g., I/O management, memory and cache management, specific data processing aspects, such as involving compression, and quantization).
Another thing to keep in mind is to sufficiently interlink the claimed features (whether abstract or tangible) to ensure operable connections and thus enhance the technical character of the invention. The reason is that only those claimed features that contribute to the technical character can support an inventive step and lead to a valid patent.
Further details can be found in our contributions to the EPO, CH, and FR chapters of the book “Artificial Intelligence and Patents: An International Perspective on Patenting AI-Related Inventions”, Kluwer Law International. Our advisors, including our lawyers for legal advice in copyright in this arena, are:
