“By Mr. Demetris Stafylas, Head os Search Directorate of the Hellenic Industrial Property Organisation (OBI).”
- Decision Τ0702/20 of the Board of Appeal of the European Patent Office (EPO).
- The appeal concerns the European patent application EP3089081 for a patent by Mitsubishi Electric Corporation, against the decision stating that the application did not contain an inventive step.
- The appellant claimed that the new and non-obvious structure of the proposed neural network model (different number of layer levels, sparse connections between nodes) ensures its technical character as it is implemented on a computer, improves learning capabilities, provides a technical advantage by avoiding overfitting, and reduces storage and computational resource requirements. It also solves the technical problem of automating human processes, and the proposed neural network model is suitable for installation on smaller devices.
- Legal framework: Art52 EPC, Art56 EPC, Key decisions G1/19, T641/00/COMVIK
- The EPO Technical Board of Appeal concluded that the difference between the invention and the prior art lies in the different layer levels of the neural network, which are connected to an error code control table. However, the specific implementation of the neural network does not require any particular computer adaptation (specific technical implementation). It agreed that computational resources and storage requirements are indeed reduced compared to fully connected neural networks in the prior art, but this cannot be regarded as a technical advantage because the proposed neural network is different, learns in a different way, and does not perform the same task as known models in the prior art. The Board also highlighted that the performance of a neural network, particularly the occurrence of overfitting, depends heavily on the characteristics of the training data. Essentially, the modified neural network model is an abstract mathematical model for solving classification or regression problems with unknown input data, and the distinguishing features(different number of layer levels – sparse connections between nodes) do not serve a technical purpose as they are not related to any specific technical implementation. Furthermore, the output of the modified neural network has no further technical application, and the appellant’s claim of automating human processes was deemed unconvincing because the training data and the technical problem to be solved were not sufficiently specified.The EPO Technical Board of Appeal rejected the appeal for the European patent application EP3089081
- According to Ministerial Decision 10374/Official Gazette ‘B 1594, 04.08.09, Article 7 “Procedure for preparing the search report or the final search report by OBI”, this decision is considered when evaluating national patent applications by the Department of Inventions, Information Technology, and Communications of the Search Directorate, Hellenic Industrial Property Organisation (OBI).”
