How to patent Artificial Intelligence inventions? Highlights of decision T 0702/20

This is illustrated by decision T 0702/20 of the Technical Board of Appeal 3.5.06. of the European Patent Office, which refused to grant a patent for a neural network with sparse connection density.

To summarize

According to the Board, a neural network defines a class of mathematical functions which, as such, is excluded from patentability. Such a network cannot therefore be patented without at least implicitly indicating a specific use case.

The disputed invention

The invention at issue in the decision concerns an adaptation of the connections between the layers of a neural network with a concept known as "loose couplings", which reduces the number of connections between the neurons of several layers of the network compared with conventional fully-connected architectures. These loose couplings are illustrated by the dotted lines in figure 2 of the patent application reproduced below.

Claim 1 relates to the ab initio architecture of the neural network, even before it is trained to perform a classification task. In particular, the loose connections between the neurons are established prior to training, independently of the training data.

Furthermore, claim 1 does not mention any specificity of the data used for training nor does it specify for what purpose the classification might be used.

Does this invention have a specific technical purpose?

• First of all, the Board points out that a neural network is made up of nodes, known as neurons, connected to each other to transmit the output of one neuron to the input of another.
  Each neuron performs a mathematical operation, for example a weighted sum of its inputs followed by a non-linear activation such as the Rectified Linear Unit function or the sigmoid function.

The architecture of a neural network, therefore, determines the operations performed by the neurons and the way in which they are connected to each other. This architecture defines a class of mathematical functions. An instance of this class is given by the set of weights that result from training the neural network to perform a given task (for example, detecting objects in images).

• The Board considers that claim 1, which relates to the ab initio architecture of the neural network even before it is trained, defines only a mathematical tool.

It thus states that this claim merely specifies abstract mathematical operations implemented by computer on unspecified data, these mathematical operations being those of defining a class of approximation functions (the network with its structure), solving a (complex) system of (non-linear) equations to obtain the parameters of the functions (learning the weights), and using it to calculate outputs for which it is not known what task they are intended.

The Board concludes that the claimed invention is not excluded from patentability because it uses a computer but  because it does not involve an inventive step.

Key takeaways

• The Board stresses that there is no doubt that neural networks can provide technical tools useful for automating human tasks or solving technical problems. However, in order to be patented, these networks must be sufficiently detailed in the application, in particular with regard to the training data and/or the technical task to be performed.
• In the present case, although the loose connections are defined by means of an error-correcting code control matrix, the applicant has not been able to demonstrate that any technical benefit would derive from the use of such a control matrix.
• Finally, the Board reminds us that it is necessary for applicants to justify their assertions, particularly when they relate to claims of technical effect.

In this case, the applicant had tried to justify that the claimed network avoided overfitting by citing a scientific article on data-driven loose coupling (in which the connectivity pattern is determined from the task to be performed, based on the distribution of training data). This article was of no help as, the Board noted that it did not relate to the claimed loose coupling where the pattern of connectivity is determined independently of the training data.

Perhaps a different conclusion would have been reached if the patent application had contained comparative examples of the claimed network with pre-existing networks, for example in terms of learning speed or accuracy of the classification obtained.

This is why we recommend that, wherever possible, such comparative examples should be included in patent applications relating to Artificial Intelligence. This is particularly important given that the question of whether post-filing published evidence should be taken into account when examining inventive step is presently debated before the Enlarged Board of Appeal with reference G 2/21.