The elusive intellectual property protection of trained machine learning models: a European perspective
Jean‐Marc Deltorn · Edward Elgar Publishing eBooks · 2022
Inference models lie at the core of artificial intelligence applications. As the end-product of the machine learning phase, an inference model combines the information derived from the training data as well as the know-how required to produce a relevant solution. Once trained, it is the inference model that is deployed in the field. It can then receive new observables as input and produce relevant results as output. The value of these objects is undeniable and their protection can be of major strategic importance for their producers. However, inference models are complex hybrid entities composed of several heterogeneous subsystems: algorithmic process, data, software, processes. At the interface between computer codes, algorithms, data structures and collections of parameters, inference models cannot be cast in a single intellectual property mould and as a consequence require examination of the applicability of the various IP rights, from copyright to trade secrets, from database to patents.