Computer-implemented inventions
At the European Patent Office (EPO), examination of computer-implemented inventions (CIIs) is well-established. A CII is defined by the EPO as one which involves use of a computer, computer network or other programmable apparatus (generally, a computer), in which one or more features are realised wholly or partly by means of a computer program.
Artificial Intelligence (AI) encompasses computers that exhibit behaviours perceived as intelligent by humans, including learning, reasoning, inferring and decision-making. Machine Learning (ML), a class of AI, gives the computer an ability to change behaviour according to experience. In other words, the computer learns.
Guidelines for Examination and Case Law of the Boards of Appeal
The EPO examines inventions based on AI and ML as CIIs. The Guidelines for Examination (revised November 2018) introduced illustrative examples for examination of AI and ML by the Examining Divisions and the Opposition Divisions. The Guidelines for Examination (revised November 2018) introduced illustrative examples for examination of AI and ML by the Examining Divisions and the Opposition Divisions. The newly-revised Guidelines for Examination (revised November 2019) provide further helpful guidance on patentability of mathematical methods while clarifying further use of state of the art terms.
Particularly, assessment of inventive step is to take into account computational efficiency of steps affecting an established technical effect of a mathematical method applied to a field of technology and/or adapted to a specific technical implementation (G-II, 3.3). Further, while terms such as “support vector machine”, “reasoning engine” or “neural network” were previously considered devoid of technical character per se, these terms are to be considered in context (G-II, 3.3.1). While the Case Law of the Boards of Appeal is, as yet, limited regarding AI and ML, the extensive corpus of case law on CIIs is expected to apply similarly.
EPO scoping workshop
The “Patenting Artificial Intelligence” scoping workshop, hosted by the EPO in The Hague on 15 October 2019, presented an insightful review of examination at the EPO and explored eight examples, inspired by previous patent applications.
Technical character
As set out broadly in the Guidelines for Examination (G-II, 3.3), AI and ML method steps may contribute to the technical character of the claimed subject matter according to two ‘dimensions’:
- By being adapted to a specific technical implementation; and/or
- By application to a field of (particular) technology.
Comment: The EPO presents these two ‘dimensions’ orthogonally, suggesting that while the claimed subject matter may include respective features of both ‘dimensions’, these respective features are mutually independent. Hence, unless a synergy between these features is identified (preferably in the description), the problem-solution approach to inventive step would be decomposed into partial problems (Guidelines for Examination G-VII, 6 & 7).
Claimed subject matter is directed to a specific technical implementation if the claimed algorithm is specifically adapted for that implementation and/or the design of the claimed algorithm is motivated by technical considerations of the internal functioning of the computer. An example is a specific ML algorithm implemented on specific computing hardware including a CPU and a GPU.
Claimed subject matter is directed to an application in a field of particular technology if a technical problem is solved and the claimed subject matter is functionally limited to a particular technical purpose. Examples are more familiar, including in the Guidelines for Examination (G-II, 3.3.1), for example image analysis, speech processing (but not text processing), medtech and autonomous vehicles.
Skilled person
The EPO acknowledges that patent applications related to AI and ML are likely to be multi-disciplinary, involving scientists, engineers and/or computer scientists. Hence, the skilled person is correspondingly likely to be a similarly-composed team. The common general knowledge of the skilled person in this field includes pre-processing data, setting parameters such as for training and selecting validation data, amongst other routine work and experimentation.
Comment: When the claimed subject matter is directed to an application in a field of particular technology, contribution from the computer scientists and/or data scientists should be included in the patent application, at least to fulfil the requirements of sufficiency and/or plausibility, as discussed below.
Sufficiency and plausibility
Notably, sufficiency and plausibility for inventions related to AI and ML were considered by the EPO.
Sufficiency relates to reproducibility of the claimed subject matter, in view of the description and the common general knowledge at the priority date. If a desired technical effect is expressed in the claimed subject matter but lacks reproducibility, an objection under Article 83 EPC should be raised. However, if a desired technical effect is not expressed in the claimed subject matter but is part of the objective technical problem to be solved, an objection under Article 56 EPC should be raised.
Plausibility relates to whether a technical effect is achieved across the whole scope of the claimed subject matter and is thus relevant to Article 56 EPC. Specific considerations noted by the include whether there is a causal link between the input and the output of the ML process and whether a trained model will produce reliable predictions for new test data.
For both sufficiency and plausibility, the EPO has indicated that experimental data and/or comparative tests may be persuasive, noting the usual conditions under which particularly post-published data may be taken into account.
Comment: The observations from the EPO regarding sufficiency and plausibility support the analysis of our previous commentary: https://www.appleyardlees.com/artificial-intelligence-and-machine-learning-sufficiency-and-plausibility/
Especially, the reproducibility checklist (after Joelle Pineau 2018), including experimental data, appears to address many of the potential issues regarding sufficiency and/or plausibility for patent applications related to AI and ML.
Examples
The eight examples explored were drawn variously from Information and Communication Technology (ICT), Mechanics and Mechatronics (MM) and Health, Biotechnology & Chemistry (HBC). In groups, each delegate considered two examples, and the summarised outcomes shared. Examples 1 and 2 are reviewed below – please get in touch to discuss Examples 3 – 8.
Example 1 – Drop out (ICT)
Drop-out is a simple training method that prevents neural networks from over-fitting; a notorious problem in ML, when a model loses its generalisation power and thus specialises too much on a given data set. Neurons are probabilistically silenced during training and the mean network is used for inference. This is computationally inexpensive and gives improvements on most benchmark tasks.
The prior art was given to be a notorious computer.
- A computer-implemented method of training a neural network including neurons, each neuron being associated with weights and a respective probability of being disabled, wherein the method comprises:
- obtaining training inputs;
- for each training input, selecting one or more neurons based on their respective probability;
- disabling the selected neurons;
- processing the training input with the neural network to generate a predicted output;
- adjusting the weights based on the predicted output.
2. The method of claim 1, further comprising: multiplying the adjusted weight associated with each neuron by the respective probability.
Example 1 falls under ‘dimension’ 1, since the claimed subject matter is directed towards a specific technical implementation.
The consistent opinion of the delegates was that examination according to the Guidelines for Examination is too narrow. Particularly, the current practice of the EPO requires limitation of the subject matter to specific computing hardware. Given the general applicability of the subject matter, such limitation was considered by the delegates to be too restrictive.
Example 2 – Industrial process (HBC)
Example 2 relates to a method for coating a workpiece by means of thermal spray coating, which is highly sensitive to fluctuations in the thermal spray. Using a CCD camera, properties (luminance distribution, particle temperature, velocity and/or size) of the thermal spray are determined.
The prior art was given to be D1: a thermal plasma spraying process in which the “in-flight” parameters are monitored and adapted using a neural network; and D2: a combination of a neural network with a neuro-fuzzy regulator.
1. A method for coating a workpiece comprising,
a) applying a material to the workpiece by thermal spray coating,
b) monitoring the thermal spray coating process on-line by detecting properties of particles in a spray jet and supplying the properties as actual values,
c) comparing the actual values or characteristic quantities derived therefrom with target values,
where there is a deviation between the actual values or the characteristic quantities and pre-specified target values,
d) adjusting process parameters for the thermal spray coating automatically by a regulator based on at least one neuronal network, wherein the regulator is a neuro-fuzzy regulator,
e) combining the at least one neuronal network and fuzzy logic rules, and thus mapping statistical relationships between input variables and output variables of the neuro-fuzzy regulator.
2. The method according to claim 1, wherein the properties detected for particles in the spray jet include particle temperature and/or particle velocity and/or particle size and/or a luminous intensity of the particles.
3. The method according to claim 1, wherein the neuronal network comprises at least four layers each having multiple neurons, wherein the neurons of an input layer map a fuzzification, the neurons of an output layer map a defuzzification, and the neurons of the layers arranged between the input layer and the output layer map a fuzzy inference.
Example 2 falls under ‘dimension’ 2, since the claimed subject matter is directed towards an application in a field of (particular) technology.
In view of the multi-disciplinary claimed subject matter, the EPO confirmed that such applications are examined by a team of Examiners, including both materials scientists and computer scientists – reflective of the skilled person. Additionally, provision (whether by inclusion in the original application and/or conditionally, after filing) of experimental data is considered beneficial, if not essential.
Comment: Generally, claimed subject matter falling under ‘dimension’ 2 is likely to involve known AI and ML, which may contribute at most to novelty, rather than inventive step. This, however, appears to conflict with the decision to grant in EP 2 773 848 B1, by way of example only, which is concerned with a problem of a drill string becoming stuck in a borehole during drilling. Granted claim 1 determines a likelihood of a drill becoming stuck, relying on four distinct machine-learning algorithms (a neural network, a decision tree, a support vector machine and Bayesian methods). The closest prior art, 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. For more detail, see:
Workshop conclusions
The extensive revisions to The Guidelines for Examination (revised November 2018, clarified November 2019) have been welcomed by applicants and patent attorneys alike, providing clearer description of practice at the EPO. Thus, the assessments of patentability of the examples by the delegates were generally accurate, even if not the desired results. That is, predictability of patentability of inventions related to AI and ML has been enhanced.
