Tag: machine learning

Semi-automatic answering/3M INNOVATIVE PROPERTIES – T 0755/18 – 11 December 2020

This decision is about the output of a machine learning algorithm. The output of the algorithm is more accurate here compared to the prior art. However, this is not a reason that the output automatically serves a technical effect. The output therefore does not automatically lead to non-technical features making a technical contribution via the output.

Object of the Invention:

  • the present application is concerned with the generation of billing codes to be used in medical billing, wherein billings are provided to an insurer for reimbursement
  • computer-based support systems have been developed to guide human coders through the process of generating billing codes
  • claim 1 specifies a computer-implemented method for improving the accuracy of automatically generated billing codes

Board I (inventive step):

  • a billing code is non-technical administrative data
  • generating a billing code is a cognitive task

Appellant (inventive step):

  • use of machine learning techniques to improve the accuracy of the machine output
  • invention is technical because it improved the system so that it would generate more accurate billing codes in the future

Board II (inventive step):

  • if neither the output of a learning-machine computer program nor the machine output’s accuracy contributes to a technical effect, an improvement of the machine achieved automatically through supervised learning for producing a more accurate output is not in itself a technical effect
  • in this case, the learning machine’s output is a billing code, which is non-technical administrative data
  • the accuracy of the billing code refers to “administrative accuracy” regarding, for example, whether the billing code is consistent with information represented by a spoken audio stream or a draft transcript
  • the learning machine to generate more accurate billing codes or, equivalently, improving the accuracy of the billing codes generated by the system, is as such not a technical effect.

Conclusion

Furthermore, the below figure shows according to G 1/19, point 85 and 86 how and when “technical effects” or “technical interactions” based on inter alia non-technical features may occur in the context of a computer-implemented process (the arrows in the figure above represent interactions and not abstract data). In this decision T 755/18 it was discussed whether the non-technical features contribute to the technical character of the invention via the output side and also via the technical implementation (although the latter is not discussed here in this commentary).

Equivalent Aortic Pressure/ARC SEIBERSDORF – T 0161/18 – 12 May 2020

This decision concerns an invention involving machine learning. The Board of Appeal ruled on the need to disclose training data in a patent application.

Object of the Invention:

  • use of an artificial neural network to transform the blood pressure curve measured at the periphery into the equivalent aortic pressure
  • Claim 1 differs from the closest prior art (general purpose computer) in that the transformation of the blood pressure curve measured at the periphery into the equivalent aortic pressure is carried out with the aid of an artificial neural network whose weighting values are determined by learning’

Board I (sufficiency of disclosure (Article 83 EPC)):

  • the application does not disclose which input data are suitable for training the artificial neural network according to the invention, or at least one data set suitable for solving the present technical problem
  • the training of the artificial neural network can therefore not be reworked by the person skilled in the art and the person skilled in the art can therefore not carry out the invention
  • no sufficient disclosure, since the training according to the invention cannot be carried out due to a lack of corresponding disclosure

Appellant (inventive step):

  • the use of an artificial neural network has the technical effect that the cardiac output can be determined reliably and precisely, while keeping the computing effort within reasonable limits, which enables integration into a mobile and correspondingly handy device

Board II (inventive step):

  • neither the claim nor the description contain details regarding the training of the neural network
  • the claimed neural network is therefore not adapted for the specific claimed application
  • the claimed effect is not achieved in the claimed method over the entire claimed range
  • the effect cannot therefore be considered as an improvement over the prior art when assessing the inventive step
  • the object is to provide an alternative to the method disclosed in D1
  • the use of a neural networks not only corresponds to a general trend in technology, it was also already known for the transformation of the blood pressure curve measured at the periphery into the equivalent aortic pressure
  • the subject-matter of claim 1 was therefore suggested to the skilled person by combining the teaching of D1 with his general technical knowledge or with the teaching of D8 –> no inventive step

Conclusion

  • in this decision, the distinguishing feature is based on machine learning (ML)
  • ML is data-driven and, therefore, the success of an ML invention will largely depend on the data on which it is trained
  • if there is too little suitable training data, it may not work
  • the application does not disclose any input data or data set
  • for AI patent applications, at least some effort should be made to explain:
    • what training data is used, and
    • why enough of it is available to train the ML system appropriately

Classifying resources using a deep network/GOOGLE – T 0874/19 – 6 July, 2022

In this decision, the EPO’s Board of Appeal clarifies that the (previously important) decision T 1227/05 is no longer followed due to the decision of the Enlarged Board of Appeal G1/19.

Object of the Invention

  • classifying search engine resources (e.g. a blog) as a spam resource or not a spam resource
  • subject matter of the invention differs from the closest prior art in that a classifier is configured to generate a respective category score for e.g. a “spam” and “not spam” category, wherein each of the respective category scores measure a predicted likelihood that a resource is a “spam” resource

Appellant:

  • following decision T 1227/05, the claimed system could at least be regarded as simulating a hardware circuit that classifies inputs and thus had a technical purpose

Board:

  • decision T 1227/05 cannot be followed as argued by the appellant in view of recent decision G 1/19, Reasons 133
  • absence of any technical effect beyond its straightforward implementation in one or more computers, the subject-matter does not involve an inventive step