With appeal cases specific to machine learning inventions still rare, it is often a good idea to look at the wider case law around mathematical methods to see what we can learn. These two cases do not relate specifically to machine learning (although the description tells us that machine learning may be used). But T1560/19 and T1561/19 provide some clear lessons on how to draft claims for the EPO and the Boards’ view of black box claim drafting. This is highly relevant to some machine learning inventions and emphasises the need to focus on the technical core of the invention. The second case also reminds us that merely reciting more specific hardware by itself is unlikely to sway the EPO.
As explained by the Board, the applications relate to monitoring a technical equipment's "behaviour". The proposed solution does so by defining numerical transfer functions to provide a mapping of input to output values. These numerical transfer functions are programmed into a multi-processor device or a multi-core processor. The mapping can be used to characterise the equipment's behaviour. During this characterisation, the equipment can operate under different behaviour conditions (“configurations”). When the mapping is determined using a model based on historical data, it can identify "trends" indicating a failure in the equipment's functioning. One example in the description (if not the claims) of how the mapping can be found is using “machine learning”.
A major issue found by the Board was whether embodiments of the description as filed should be considered when construing claims. In this respect, the Board noted that Article 84 EPC requires the claims to define the subject matter for which patent protection is sought and that the meaning of the features of the claims should therefore be clear for the person skilled in the art from the wording of the claims alone (see e.g. G 1/04, Reasons 6.2). In other words, given that claims of a patent application are typically directed to a person skilled in the field of the application, the claims should essentially be read and interpreted by such a skilled reader on their own merits, rather than with the aid of the description and drawings. This is a good reminder of best practice for EPO claims, which must be understood by themselves to be clear. While this gets drilled into us in patent school in Europe, practice differs on this point in the US, where the patent specification can be its own dictionary much more readily than in Europe. So, this is an essential practice point for applications crossing the Atlantic in an easterly direction.
For the application at hand, the Board noted that, for example, the clause "for characterising behaviour properties of equipment under observation" is unclear because it is not apparent how the equipment's behaviour properties are characterised by the data processing device by whom or by which entity the equipment is under observation. This was because no details are provided in claim 1 about the relationship between the "equipment's behaviour properties" and the "configuration", other than the vague indication that some output values represent the behaviour properties. The "configuration" merely specified an "input-to-output mapping" that is performed by the processing units. Moreover, the "input" of the "input-to-output mapping" was completely arbitrary. In particular, the claimed "data processing device" did not include any features that would enable an input based on an observation. Clearly, while the Boards of Appeal have often emphasised that broad claims are not necessarily unclear, there is a line where a claim becomes so vague that it cannot anymore be understood at a technical level. For the Board, this case crossed that line, and the fact that more information was provided in the description did not help.
While the Board could have stopped there, they went on to consider inventive step, in the process providing us with another reminder of drafting best practices. The appellant tried to argue a technical effect on the basis that the skilled person would understand that the claim features are characterised by pattern recognition performed on the behaviour of master equipment and that after a training period, the claimed data processing device would have been able to achieve an effect of "re-recognising certain signal patterns”. The Board did not buy it and believed that this effect is not credibly achieved by the claimed features. They noted that:
-
Instead, the claim used general terms such as "numerical transfer functions", "input-to-output mapping", and "configuration", which are inherent to any processing unit like one of those mentioned in the claim.
-
No other technical effect could be recognised that could credibly be brought about by the claimed features because, in fact, the claimed data processing device acted as a "black box" using abstract mathematical method steps, without achieving any credible technical effect.
The Board recalled that, according to Reasons 124 of G 1/19, "only those technical effects that are at least implied in the claims should be considered in the assessment of inventive step". Here is another important best practice point for European claim drafting: any technical effect that one wants to rely on for inventive step has to be brought about, at least implicitly, by the features of the claims. The fact that claim features may be able to achieve a technical effect is neither here nor there. The claims must be limited with features that achieve the technical effect to be relied on.
The appellant also submitted several auxiliary requests, but these did not convince the Board either, for similar reasons. Neither did the recitation of “a plurality of processor cores” in the parallel case help. It did not help address the clarity issues and did not help show a credible technical effect. The appellant tried to argue the presence of multiple processor cores implied some sort of scheduling on the fly and saving of energy. While the Board agreed that some sort of scheduling was implied, since the claims were silent about how any scheduling would be performed, without any mention of “on the fly” or “energy saving”, the argued technical effect was also not credibly brought about by the claimed features.
Unfortunately for the appellant, the appeals were dismissed for lack of clarity and lack of inventive step of the claimed subject matter. Luckily for us, this case provides a valuable reminder and illustration of two core best practice points for claims that are to be successful at the EPO.