EPO Board of Appeal clarifies the bar for AI/ML patent disclosure: T 1669/21 case note

EPO Board of Appeal clarifies the bar for AI/ML patent disclosure: T 1669/21 case note

The EPO Board of Appeal recently handed down a decision in case T 1669/21 that has important implications for patent practitioners drafting machine learning patent applications. The case involved a patent application for a method of determining the condition of a refractory lining in a metallurgical vessel using a “Rechenmodell” ("computational model"). The patentee argued that this implied the use machine learning, but the Board ultimately disagreed, revoking the patent for lack of a sufficient disclosure that would be needed to enable the use of all possible computational models. The Board also considered what would have been required if the claims were actually limited to machine learning and in doing so provided helpful guidance for practitioners.
 

Legal Context

The decision turned on Article 83 of the European Patent Convention (EPC), which requires that a patent application disclose the invention in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art. The Board found that the patent application failed to meet this requirement, particularly concerning the machine learning aspects of the claimed invention.
 

Arguments and Findings in T 1669/21

  • Rechenmodell (Computational Model): The patentee argued that the term “Rechenmodell” in Claim 1 referred specifically to a machine learning model, pointing to the use of terms like “adapted” and “regression analysis” as indicative of machine learning concepts. They also cited Claim 5’s reference to a neural network model. However, the Board disagreed, finding that “Rechenmodell” could encompass analytical models as well, and Claim 5 did not limit the broader interpretation of Claim 1. Since the patent did not provide details for creating any kind of analytical model, the Board found it did not meet the requirements of Article 83 EPC.

  • Machine Learning Model Specificity: Even considering a machine learning model, the patent lacked crucial specifics. The applicant argued that the extensive knowledge of machine learning and the availability of machine learning libraries would enable a skilled person to implement the invention. However, the Board countered that the patent offered no guidance on selecting a specific model architecture, the mathematical modelling of nodes, or the learning procedure from the wide variety of available options. This lack of specificity presented a significant hurdle for a skilled person to implement the invention.

  • Selection and Representation of Input Parameters: The patent broadly defined types data and parameters without specifying which measurements within those categories were relevant for the model. The applicant argued that a skilled person would know which parameters were relevant and could select a technically sensible set. They asserted that machine learning would in any case filter out irrelevant inputs, so the inclusion of less relevant parameters would not impact the model's predictive power. However, the Board found that without a concrete example demonstrating successful prediction, it was not evident that the chosen parameters would be sufficient or that irrelevant parameters would not hinder the learning process. They emphasized the need for the patent to enable implementation across the claimed breadth, including guidance on selecting and representing parameters, which was lacking in this case.

  • Sufficiency of Training Data: The patent did not specify the scope of operating conditions or variations needed in the training data. The applicant suggested that data from normal operations would suffice, with the model down-weighting parameters that remained constant. However, the Board found no evidence supporting this claim or addressing the risk of learning spurious correlations from limited data. They also pointed out that training on a limited dataset would restrict the model's applicability, contradicting the broad claims.

  • Overall Enablement: The Board concluded that the patent’s disclosure did not enable a skilled person to carry out the invention across its claimed breadth. They criticized the patent for relying on the general idea of a "black box" computational model without providing sufficient details for implementation, particularly concerning parameter selection, model architecture, and training data. This lack of detail was deemed disproportionate to the breadth of the claims and the effort required for a skilled person to fill the gaps.

 

Points Decided

Here’s a breakdown of the key points decided in T 1669/21:

  • Detailed disclosure of the ML model is crucial: The Board held that simply mentioning the use of machine learning is insufficient. The patent application must provide a comprehensive description of the model's architecture, including its topology, mathematical modelling of nodes, and learning procedures. Applications must go beyond mere name-dropping and clearly explain the specific AI/ML model and how it's implemented.

  • Clear definition of input and output variables is essential: The use of broad terms to describe input and output variables is not acceptable without further detail. The patent application must clearly define the specific variables used by the AI/ML model to ensure that a skilled person can reproduce its functionality. Vague categories won't cut it; concrete examples are needed.

  • Guidance on parameter selection is required: The Board found fault with the lack of guidance on selecting specific parameters within broadly defined categories. Applications must provide clear instructions and criteria for choosing appropriate parameters. This ensures the skilled person can implement the invention across its full scope without undue burden. A working example is at least a good start.

  • Disclosure of training data quantity and quality is necessary: The Board criticized the lack of information about the training data used for the AI/ML model. Applications must address the source and characteristics of the training data, ensuring it adequately represents the relevant parameter space for the claimed functionality. Don't just assume the skilled person knows where to get the right data and what that would look like.

To summarise, the decision emphasises the need to disclose at least one working example in sufficient detail to put at least one embodiment of the claimed invention into effect. This can be the starting point for guidance of how this embodiment can be varied, to enable the desired broad claim scope. A rich disclosure, in terms of specific working embodiments and guidance as how they can be generalised, reduces the risk that a broad claim scope is not enabled. In addition, the decision confirms the approach set out in the EPO Guidelines of examination and clarifies the disclosure requirement with respect to training data: sufficient information about the training data must be provided to enable the skilled person to work the invention, disclosure of the data itself is not needed.
 

Implications for Practitioners

This decision clarifies the bar for disclosure in machine learning patent applications. Here’s what it means for practitioners:

  • Be explicit about AI/ML use: Don't rely on general terms to imply the use of AI/ML. State it clearly and provide specific details.

  • Describe the AI/ML model in detail: Provide a comprehensive explanation of the chosen model, its architecture, learning process, and implementation.

  • Define input and output variables precisely: Use concrete measurements and avoid broad categories when defining variables.

  • Offer clear guidance on parameter selection: Provide examples, criteria, and instructions to enable the skilled person to select appropriate parameters for implementing the invention.

  • Disclose relevant the training data: Specify the source, quantity, and quality of training data, and demonstrate its relevance to the claimed functionality. Disclosure of the data is not needed but example data points can help illustrate these concepts.

The Board's emphasis on detailed disclosure and enablement aligns with the increasing complexity of AI/ML technologies. This decision provides a helpful summary of the scrutiny patent applications in this field are likely to receive. Practitioners need to make sure their drafting strategies are fit for purpose to meet these requirements and ensure their clients' inventions receive adequate patent protection.


For more information, please contact Alexander Korenberg or your usual Kilburn & Strode advisor.

Let us keep you up to date. If you’d like to receive communications from us, ranging from breaking news to technical updates, thought leadership to event invitations, please let us know.

Connect with us