At the European Patent Office (EPO), the Boards of Appeal have long held a view that features of an algorithm underlying a computer-implemented method, falling under the exclusions of Article 52(2) and (3) EPC, forbidding patenting of a computer program “as such”, provide a technical contribution only to the extent that they interact with the technical subject matter of a claim for solving a technical problem (T 154/04). A technical interaction may be present if technical considerations motivating the algorithm’s design can be identified that make the algorithm suitable for being performed on a computer and that ‘go beyond merely finding a computer algorithm to carry out some procedure’ (T 1538/09). However, simply improving algorithmic efficiency has been assessed by the Boards to not provide a technical effect in support of inventiveness (T 1784/06, T 42/10, T 1370/11).

Decisions of the Boards of Appeal and granted European patents relating to use of artificial intelligence (AI) or machine learning (ML) are as yet few. However, what decisions there are do not depart from the well-established requirements for patentability of computer-implemented inventions at the EPO.

In T 2418/12, the Board of Appeal recently had to address the question of whether a ‘statistical classification and machine learning technique’ provides a non-excluded, technical contribution that merits grant of a patent. This case concerned related-term suggestion, which allows relevant keywords, for search engine result optimisation, to be identified without human intervention. Claim 1 of the Main Request recited, as a preamble, ‘A computer-implemented method for related term suggestion’ while the characterising part of this claim detailed, extensively, an algorithm for executing the method of the preamble. Claim 1 of the Main Request did not require a ‘statistical classification and machine learning technique’. In their decision, the Board considered that only the preamble was a technical means and that its computer implementation, as set out in the algorithm of the second part of the claim, was not a technical feature. That is, the algorithm of claim 1 of the Main Request fell firmly within the exclusions of the EPC preventing patenting of computer programs as such. Auxiliary Requests I, II and III suffered similar fates. However, Claim 1 of Auxiliary Request IV additionally required the algorithm to include ‘generating a trained classifier … by using a statistical classification and machine learning tool.’ While the Board accepted that this limited the choice of algorithm, the Board decided that ‘this algorithmic feature does not render the non-technical algorithm technical.’ The Appeal was dismissed, as none of the requested claim sets included non-excluded technical matter supporting an inventive step.

This decision is consistent with that of T 1510/10, in which the Board of Appeal had to consider whether using machine-learning algorithms could contribute to inventive step. This case concerned ranking information, particularly live web applications, based on interest and/or importance. The Board highlighted that the claimed subject matter failed to define any particular method of machine learning – not even one was described in the application. Rather, machine learning was presented in the application as known. Thus, the Board decided that ‘no inventive step can derive just from the use of machine learning.’ Again, the Appeal was dismissed.

In T 1285/10, however, grant of a patent was allowed following remittal to the department of first instance, but not because of the artificial intelligence routines that the claimed method used. This case related to an artificial intelligence system for genetic analysis, claiming a method for diagnosing and recommending treatment for a physiological condition. Three of the five claimed method steps related to handling hybridization data of an array of oligonucleotides of about 20 mer to about 25 mer or peptide nucleic acid probes. Particularly, step (iv) required using artificial intelligence routines to determine the most likely pathological or physiological conditions suggested by comparative analysis of hybridization profiles. However, this use of AI was not at issue before the Board of Appeal since it was common ground that use of AI generally was already known. Rather, the case turned on the type of data. While the Board was only reviewing the first instance decision on added matter and sufficiency, it did not come to a decision on on inventive step, but made an obiter dictum observation that it considered the claims of the requests to be obvious in the light of the prior art. After remittance, the department of first instance decided that the auxiliary request considered  in the Appeal was inventive, in view of its use of hybridization information from an array of peptide nucleic acid probes, and a patent granted, published as EP 1 222 602 B1.

So when considering inventions that use artificial intelligence or machine learning, wherein lies the invention? Following T 1510/10 and T 2418/12, the invention should not be in the use of AI or ML per se, as these fail to provide non-excluded technical contributions even if they do give novelty. Rather, the invention should be in the non-excluded technical solution to a technical problem – as apparent in T 1285/10 and exactly as according to the established case law for assessment of patentability of computer-implemented inventions at the EPO. To see this further, it is helpful to review other granted European patents that claim use of AI or ML.

EP 2 214 403 B1, granted before publication of the decisions of T 1510/10 and T 2418/12, is concerned with a solution for video jitter, a common problem when capturing video using a handheld device. Granted claim 1, unopposed, requires that an image stabilization software correction parameter (a parameter) is calculated, at least in part, using machine learning:

• An image processing device that calculates a parameter used in alignment performed on at least two images captured by a photographing device (102) that has a sensor (101, 111, 112) detecting a movement, said image processing device comprising:
• an image obtaining unit (101, 122) configured to obtain an image captured by the photographing device (102);
• a sensor information obtaining unit (121) configured to obtain sensor information that is an output signal from the sensor, the output signal being provided when the image obtained by said image obtaining unit is captured;
• a constraint condition generating unit (130) configured to generate a constraint condition using the sensor information obtained by said sensor information obtaining unit such that a value of a parameter to be calculated falls within a range; and
• a parameter calculating unit (140) configured to calculate the alignment parameter related to the image obtained by said image obtaining unit, according to the constraint condition generated by said constraint condition generating unit,

Wherein said constraint condition generating unit includes:

• a feature vector generating unit (131) configured to generate a feature vector that is obtained from a detected movement of the photographing device, the movement being indicated by the sensor information obtained by said sensor information obtaining unit (121); and
• a motion classifying unit configured to identify the movement of the photographing device according to the feature vector generated by said feature vector generating unit, on the basis of an association between the feature vector and the movement of the photographing device,
• the association is obtained as a result of previously-executed machine learning of the feature vector and an actual movement of the photographing device,
• said constraint condition generating unit is configured to generate the constraint condition by determining the range corresponding to the movement of the photographing device, the movement being identified by said motion classifying unit, and
• said parameter calculating unit is configured to calculate the parameter by searching for the parameter within the determined range.

Arguing in support of inventive step during Examination, the Applicant submitted that the movement of the photographing device is identified based on a result of previously-executed machine learning of a feature vector and an actual movement of the photographing device, allowing for a smaller search range to be used, increasing accuracy of the parameter (for example, a tilt angle of the image or an amount of translation) and reducing operation cost for the searching.

The effect is technical: improving image stabilization. While machine learning is a part of the claimed solution, the invention arguably is in determining the type of movement (conveniently classified by the machine learning) and hence more accurately and efficiently calculating a parameter for image stabilization.

EP 2 773 848 B1 is concerned with a problem of a drill string becoming stuck in a borehole during drilling. Granted claim 1, as yet unopposed, determines a likelihood of a drill becoming stuck, relying on four distinct machine-learning algorithms:

1. A method comprising:

• receiving, at a computer system, a plurality of drilling parameters from a drilling operation;
• applying, by the computer system, the plurality of drilling parameters to an ensemble prediction model comprising at least three machine-learning algorithms operated in parallel, each machine-learning algorithm predicting a probability of occurrence of a future stuck pipe event based on at least one of the plurality of drilling parameters, the ensemble prediction model creates a combined probability based on the probability of occurrence of the future stuck pipe event of each machine-learning algorithm; and
• providing an indication of a likelihood of a future stuck pipe event to a drilling operator, the indication based on the combined probability,
• wherein applying the plurality of drilling parameters to the ensemble prediction model further comprises:
• applying at least a portion of the plurality of drilling parameters to a first machine-learning algorithm, the first machine-learning algorithm predicting a first probability of occurrence of the future stuck pipe event;
• applying at least a portion of the plurality of drilling parameters to a second machine-learning algorithm distinct from the first machine-learning algorithm, the second machine-learning algorithm predicting a second probability of occurrence of the future stuck pipe event;
• applying at least a portion of the plurality of drilling parameters to a third machine-learning algorithm distinct from the first and second machine-learning algorithms, the third machine-learning algorithm predicting a third probability of occurrence of the future stuck pipe event;
• applying at least a portion of the plurality of drilling parameters to a fourth machine-learning algorithm distinct from the first, second and third machine-learning algorithms, the fourth machine-learning algorithm predicting a fourth probability of occurrence of the future stock pipe event; and
• combining, by the computer system, the first, second, third and fourth probabilities, the combining creating the combined probability,
• wherein the first to fourth machine-learning algorithms respectively are a neural network, a decision tree, a support vector machine and Bayesian methods.

It was accepted in Examination that the closest prior art, which is concerned with the same problem, teaches use of echo state networks for stuck pipe detection. An echo state network is a recurrent neural network with a sparsely connected hidden layer and hence a machine-learning algorithm. According to the Applicant during Examination, combining the claimed distinct machine-learning algorithms may increase the accuracy and thus decrease false positive indications.

Intention to grant predates the decision of T 2418/12 by a month. Yet in view of T 2418/12 and reinforced by T 1510/10, grant now of such a claim  would seem contrary to at least these decisions. The difference over the closest prior art is combining a known neural network with a decision tree, a support vector machine and Bayesian methods. If one machine-learning algorithm cannot provide a non-excluded technical contribution and hence confer an inventive step, even a novel combination of distinct multiple-learning algorithm cannot either. Furthermore, if the increased accuracy is due simply to improving algorithmic efficiency by this combination, this cannot be a non-excluded technical effect (T 1784/06, T 42/10, T 1370/11). Alternatively, it could be argued that the effect, achieved by the specific combination, of decreasing false positive indications, somehow represents a real technical effect and hence is determinative.

For applicants, patentees and opponents alike, these decisions help in assessing whether a claimed invention, that uses AI or ML, will succeed or fail. AI or ML may bring novelty, but AI or ML per se does not give the non-excluded technical contribution necessary for an inventive step. At minimum, the claimed invention should provide a technical effect in the real world that is more than simply a way of speeding up arriving at the solution to a problem. Achieving the technical effect should not be reliant solely on the AI or ML, both being arguably well-known and having expected outcomes, at least in a sense of providing improvements. Nevertheless, where AI or ML is used, the application should comprehensively describe the detailed implementation, functionally and structurally – even if only for sufficiency or a view towards the US patent office requirements for patentability.

At Appleyard Lees we have a large and experienced team of attorneys who combine a broad range of legal and technological expertise with real world commercial insight.  If you need help with navigating the complex world of intellectual property, contact us here.