EP 3 660 397 B1 relates to a device, a system, a program, and a method for estimating the composition of waste in a waste treatment plant.
Claim 1 as granted defines a device comprising a “training data generation unit“, a “model construction unit” and an “estimation unit“. The device is for training and using machine learning for obtaining a “value representing the composition of waste” in a waste pit upon inputting “data of” a captured image of the waste.
Brief outline of the procedure
The opposition was rejected and the opponent appealed.
The board decided that the invention was insufficiently disclosed, and the patent could not be maintained as granted as well as to according AR1-3. AR4-7 were not admitted under Art 13(2) RPBA.
The conditions for maintenance according to AR8 were not met.
The patent was thus revoked.
The proprietor’s point of view
The proprietor submitted that the description disclosed a limiting definition that “the value representing composition of waste is an index showing the degree of flammability or non-flammability of waste.
The proprietor also submitted that claim 1 did not require a specific level of accuracy. Any value that estimated the composition as being “better than arbitrary” could be considered to “represent” the composition as defined in claim 1.
The proprietor stated that the patent in fact disclosed a specific example of carrying out the invention.
The proprietor referred to well-known properties of waste and how to classify it as done by a human operator on the basis of captured image data. The skilled person was also aware from their CGK that emulating this by machine learning represented “one way” of carrying out the claimed invention.
Accordingly, taking into account the CGK of the skilled person, the patent sufficiently disclosed one way of carrying out the invention.
The proprietor also submitted “Annex 1” setting out some details as evidence that the invention could be carried out.
The opponent’s point of view
The opponent submitted that the invention of claim 1, which concerned machine learning, did not meet the corresponding requirements set out in T 161/18 and T 1669/21 for sufficiency of disclosure of machine learning inventions.
The board’s decision
The board noted that a certain number of the features claimed, like “data of” a captured image, value representing composition and accuracy, were not limiting the subject-matter of claim 1.
“Machine learning model”
Claim 1 defines a “model construction unit” adapted to “construct a model by performing learning using the training data”, but it does not explicitly limit the type and architecture of the model, nor the learning method.
The description does not provide further details in this regard either. The relation disclosed in the description between the input x and the output y and the internal parameter adjusted by machine learning applies to virtually any machine learning approach, and the list at the end of the corresponding passge is an almost arbitrary enumeration of well-established, but very different machine learning techniques.
The patent does not provide any further details on the learning method. Accordingly, the type, implementation and training of a machine learning model suitable for the claimed purpose remain open.
With regard to the decisions referred to by the opponent, in the case underlying T 161/18 the competent board concluded that the disclosure regarding the training of the artificial neural network in question was insufficient. In T 1669/21, the competent board held that the patent did not disclose any specific example for carrying out the invention, which specific machine learning model and combination of specific input values could be used for predicting the claimed output parameter, and how the invention could be carried out over the whole claimed breadth.
In the present board’s view, these aspects, in that particular case, illustrate the glaring gap between the breadth of the claimed invention and the level of detail in the patent, and do not constitute generally applicable criteria required for sufficiently disclosing a machine learning invention.
The board noted that the patent does not contain a specific example of the claimed invention. That is, it does not disclose any specific combination of certain “data of” captured images and a particular “value representing composition” of the waste in the images, nor does it provide any details on the implementation and training of an exemplary machine learning model, or any information on the achieved accuracy of estimation. The patent does not contain any concrete, reproducible example of implementation of the invention.
A specific example is not in itself an absolute requirement for sufficient disclosure, provided that the skilled person is aware of “at least one way” of carrying out the invention in other ways, for example, through the generic disclosure in the patent or the common general knowledge.
In the present case, providing such an example could have demonstrated that the invention is workable at all, at least in this specific case of the example. It could have served as a reference to better understand the claimed invention, its terms and purpose and the achievable or expected level of accuracy.
The board held that the example in “Annex 1” was not sufficient for establishing how the invention can be carried out over the whole breadth of claim 1, and its admittance under Art 13(2) RPBA, was not decisive and could be left open.
T 149/21, Reasons 3.2, reflects the general principle that the protection obtained with the patent has to be commensurate with the disclosed teaching.
The patent in suit teaches the general idea of using machine learning to infer properties of the waste composition that could be relevant for operating and controlling a waste incineration plant from images of the surface of the waste pit.
As the disclosure is mostly limited to stating a “result to be achieved”, the mere idea does not enable the skilled person to carry it out.
The task that the skilled person faces is to design solutions using machine learning models for all conceivable input data and output “values representing composition”.
Each evaluation represents a considerable effort in itself. Exploring all the possible combinations of the parameters would require a comprehensive research programme and would place an undue burden on the skilled person.
The undue burden is due to the fact that the skilled person has insufficient information on the relevant criteria for finding workable solutions across the whole breadth of the claim.
The undue burden also does not stem from the fact that each evaluation alone requires some effort. Instead, it is due to the fact that the skilled person has no guidance from the patent on how to choose a suitable combination from across the claimed breadth to start from.
All of this reflects the fact that the disclosure in the patent is not proportionate to the breadth of the claim, so that the protection obtained is not commensurate with the disclosed teaching.
Comments
Although the patent is classified in F23G 5/50, F23G 5/44 and B65F 5/00, dealing with incineration and transport of waste, the present decision has been taken by Board 3.5.06 in an enlarged composition with 3 TQM and 2 LQM.
That machine learning or the use of AI and LLM will make inroads in lots of technical domains far distant from IT as such, is to be expected. However, the present decision should be a warning to applicants/proprietor when introducing machine learning, AI or LLM, in their applications/patents.
Although examples are not necessarily required for sufficiency, in a case like the present one, it appears a necessity. And then only one example will have been disclosed, but the disclosure should encompass the whole claim breadth.
In spite of the classification of the subject-matter of the present patent, the ED should have comprised one member of the directorate dealing with machine learning, AI and LLM, even if this might slow down the production.
The ED in charge never queried in the slightest the problems related with the machine learning. It was most probably not aware of those, which is acceptable as such, but should have sought help on those aspects it does not mastered.
T 1669/21 was commented in the present blog under “The magic effect of a neuronal network in determining the wear and tear of a metallurgic vessel”.
The present patent is of the same kind, as it merely defines a result to achieve.
Comments
Leave a comment