T 0183/21 - Feedback control of model training

T 0183/21 - Feedback control of model training

This recent decision concerned a recommender system for recommending items to users, for example video in a VOD system or products. The invention was concerned with how to control the performance of the recommender system. While recommending products is not considered to be technical by the EPO, the way performance was controlled was found to be new and involved an inventive step. The case is an interesting illustration of how a technical implementation of an otherwise non-technical invention can lead to a patent being granted and is a rare example of the Boards of Appeal allowing an appeal in a case relating to machine learning.   

The claimed recommender system used feedback control of the system’s performance as follows. A level of performance of the recommender system is measured and compared to a desired level of performance by a feedback controller. The feedback controller increases or decreases the amount of training data to be used for update the recommender system based on the comparison (specifically using a PID controller with the level of performance as the controlled variable and the amount of training data as control variable). This meant that an amount of training data was used that was in line with, but not more than, what was needed to achieve the desired level of performance. As a consequence, network bandwidth and compute resource requirements were reduced as compared to a system that would obtain all available training data for the best possible performance.  

The Board noted that recommending products is not recognised as having technical character. However, considering the closest prior art document, the Board found that the claimed subject matter was distinguished by the feedback loop that controls the amount of training data used. The objective technical problem to be solved was therefore to reduce the use of network bandwidth and the amount of storage in a communications system including a client device and a recommender system in communication with the client device. The closest prior art did not have this feedback mechanism and indeed focussed on optimal performance of the recommender system, rather than controlling the system towards a desired, not necessarily optimal, level of performance to limit use of resources. Therefore, an inventive step was acknowledge for the claimed subject matter and the application was remitted to the first instance for further prosecution in respect of the dependent claims.  

The decision is of interest, since there was no question that there is indeed a technical effect achieved by technical means in the form of the feedback loop. This point was not expanded on in the decision and, indeed, was taken for granted, although the first instance division denied any such effect existed. Rather, the discussion appears to have focussed on whether the effect applied to the whole scope of the claims, that is whether all essential features needed to achieve the effect were recited.   

While the decision is light (to put it mildly) on the reasons for acknowledging the technical effect, the case is a useful example of an effect being acknowledged for a technical implementation of an otherwise non-technical method, that is one motivated by technical considerations of the functioning of the technical system implementing the non-technical method.  

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

Cookies improve the way our website works. By using this website you are agreeing to our use of cookies. For more information see our cookie policy I accept