Artificial intelligence (AI) patents in Japan reflect a
pro-patent policy as a result of the fourth industrial
revolution. In the machine learning field, examiners at the
Japanese Patent Office (JPO) reported that the number of patent
applications has increased by 78% per year and the patent grant
rate has reached 90%. In addition, the JPO made it clear that
claims of trained classifiers (trained models) could be
In this article, we propose adding extended intelligence
patents (EI patents) to your patent portfolios. EI patents
relate to the human creation of technical ideas extended by AI.
EI patents are typically related to the behaviour of products.
The behaviour is based on outputs of trained classifiers. The
behaviour directly solves business and technical problems in
various situations (see Figure 1). The behaviour could be
implemented using AI or without AI.
We will discuss an outline of EI patents, trained
classifiers and IF patents and an example of an EI patent.
Those who are not familiar with AI technology are encouraged to
read the section on the simplest example of an EI patent.
EI patents can often be better conceived by product experts
than AI specialists.
Figure 1: Extended intelligence patents
Figure 2: Third step and EI patents
|AI system process
1) Training step
AI is training datasets.
2) Runtime step
AI output to inputs.
3) Solution step
AI controls products which resolve
|AI patent at each step
Patent related to behaviour of products with
outputs at "2) runtime step".
AI runtime patent
Patents related to input/output device with
classifier trained at "1) Training step".
AI training patents
Patents related to training devices.
Figure 3.1: AI training patent at the (1) training
Figure 3.2: AI runtime patent at the (2) runtime step
Figure 3.3: EI patents at the (3) solution step
Outline of EI patents
AI systems seem to have only a training step and a runtime
step when it comes to patents. However, we suggest a solution
step as a third step (see Figure 2).
EI patents are AI patents at the solution step. They are
related to the behaviour of products, structures of products,
and materials that are based on the outputs of trained
classifiers. The behaviour is movement, control, prediction, or
graphical user interface (GUI) output. For example, AI performs
a control which causes products to perform a behaviour.
Examples of behaviour
In the cases in the grid entitled "Examples of behaviour",
the subject of each claim of the EI patent is the behaviour of
the car, the robot arm, the production machinery, or the
trading device, respectively.
In another example, when AI predicts more suitable
structures and materials for laser radiators or control chips,
EI patents are patents for the structures and the materials
predicted by AI. In the final section of this article, we will
introduce the simplest example of EI patents.
Effectiveness of EI patents
We will compare patents at each step in light of patent
EI patents have two advantages for the patent holder: it is
difficult to avoid EI patent infringement and it is easy to
discover EI patent infringement.
AI training patent infringement
AI training patents are patents at the training step.
AI training patents are related to the gathering, selecting,
and converting of training datasets or reward signals.
Alternatively, they are related to selecting and setting
classifiers, or new classifiers. It is possible to enforce AI
training patents against AI systems at the training step.
Figure 4: Training step and runtime step
However, it is impossible to enforce AI training patents
against non-AI systems because non-AI systems are developed by
programming, but are not trained. It is impossible to enforce
the patent of one type of model against other kinds of models
(see Figure 3.1). Furthermore, it is difficult to discover AI
training patent infringement in some cases because, apart from
at the runtime step, AI training can be conducted
AI runtime patent infringement
AI runtime patents are patents at the runtime step.
AI runtime patents are related to input/output devices with
trained classifiers. It is possible to enforce AI runtime
patents against AI systems at the runtime step.
However, it is impossible to enforce AI runtime patents
against non-AI systems because non-AI systems run with
application programming interfaces (APIs), databases, and
functions, but not with trained classifiers (see Figure
EI patent infringement
EI patents are patents at the solution step.
It is possible to enforce EI patents against both AI systems
and non-AI systems because EI patents are related to the
behaviour of products. It is also difficult to avoid EI patent
infringement because the behaviour which can resolve problems
is the same regardless of whether it is with trained
classifiers or with API/database/functions. Furthermore, it is
easy to discover patent infringement since the behaviour of
products is directly observed (see Figure 3.3).
EI patents are very important patents. Prior to explaining
EI patents in detail, we will introduce Japanese AI
Trained classifiers and IF patents
A long time ago, in March 2017, the JPO presented claims of
trained classifiers in the Examination Handbook. Patents for
trained classifiers were a hot topic. They are patents at the
training step and runtime step. In these steps, we will
introduce claims of a process for producing trained classifiers
and interface patents (IF patents) for AI (see Figure 4).
Figure 5.1: Example of IF patent at the training step
Figure 5.2: Example of IF patent at runtime step
Figure 5.3: Example of production methods of trained
Patents related to trained classifiers
Claims for trained classifiers
The Handbook contains the case example, "Case 2-14 Trained
Classifier to Analyse Reputations of Accommodations". Through
this example, the JPO made it clear that the claims of trained
classifiers could be patented. The JPO interpreted trained
classifiers to be a "computer program, etc." (Japanese Patent
Law, Article 2, Paragraph 3, Number 1). Thus, the inventions of
trained classifiers can fall under an invention and be
Claims of a process for producing a trained classifier
Along with proposing claims of trained classifiers, we
additionally suggest claims of a process for producing trained
classifiers. Firstly, the patents of the suggested claims can
protect both the process and the product (trained classifiers)
in Japan. Secondly, we noticed that the claims for trained
classifiers can be written as product-by-process (PBP) claims.
We may need to consider avoiding the issue of PBP claims.
Thirdly, we found a Japanese patent (patent number 6216024)
which has claims of a process for producing trained
IF patents with an example
IF patents are input/output patents without inner structures
that are for example, the topologies of neural networks setting
each layer of deep learning. In other words, there is no inner
structure of a classifier in their claims. IF patents at the
training step are related to training datasets.
IF patents at the runtime step are related to input/output
devices with a trained classifier.
IF patents are important patents because they can be
enforced against not only the same kind of classifiers but also
other kinds of classifiers. It is a common function of training
to minimise the errors between training datasets and
input/output with training classifiers.
The first example of a claim (Japanese Patent Number
6063016) as an IF patent at the training step is shown in
Training datasets are status variables and the results of
judging in the example.
There are training datasets but no classifiers in the claim.
Thus, the patents of the claims can enforce the right against
any kind of classifier and any inner structures of a
If you would like to include distillation models in the
scope of the patent right, it is best to change "a status
observing part observing … as status variables" to "a
status acquiring part acquiring … as status
The second example of a possible Japanese claim as an IF
patent at the runtime step is shown in Figure 5.2.
Inputs are status variables, and outputs are correction of
the operation command in the example.
There are inputs/outputs and the classifier, but no inner
structures of the classifier in the claim. Thus, the patents of
the claims can enforce the right against any kinds of
classifiers and any inner structures of a classifier too.
In the case of supervised learning, according to the above
claim of IF patent at the training step (claim 1 of Japanese
patent number 6063016), the outputs from the classifier are
considered to be the probabilities of whether or not an
abnormality will occur in the electric motor. In this case, it
is possible to change "the classifier which…according to
claim 1" to "the classifier which outputs information
indicating whether or not an abnormality will occur in the
electric motor for the status variables inputted" in the above
claim of IF patent at the runtime step.
Production method of trained classifier of IF patents
We would suggest adding the invention category production
methods of trained classifiers. There are no inner structures
of classifiers in the claims of IF patents (see Figure
We also propose claims of trained classifiers, but it is
necessary to pay attention to the fact that there are inner
structures of a classifier (two neural networks) in the JPO's
Whether the internal structure should exist in the claim or
not depends on future practices and application contents.
Simplest example of EI patent
Behaviour of a robot arm
There is an interesting video showing a robot arm controlled
by AI. The robot arm picks up cylinders similar to bins. The
video shows the success rate of picking up the cylinders. After
AI trained 5,000 samples, the success rate dramatically
increased from 60% to 90%. We must pay attention to the result
of the analysis. The result is that "round-shaped facets are
actually easy targets" (see Figure 6 or video) after
As the simplest example of EI patents, we propose the claim
in Figure 6.2 based on the result of the analysis. The claim is
directed to a robot arm, and there is no description about AI
at all in this claim.
Figure. 6.1: Behaviour of robot arm
Figure 6.2: Simplest example of an EI patent
Advantages of EI patents
The claim in Figure 6.2 of an EI patent shows the following
Firstly, in the claim, the structure of AI systems at the
training step and runtime step was not described at all. In
other words, it is possible to enforce this EI patent against a
robot arm performing the above behaviour due to programming in
non-AI systems. Secondly, it is also possible to keep training
methods and inner structures of a classifier confidential in AI
systems because it is not necessary to describe them in the
patent application. Thirdly, the patent practices for EI
patents are as usual since the claims of EI patents are the
same as claims which used to be described in non-AI systems.
See Figure 7 for advantages of AI patents.
Advantages of EI patents
Figure 7: Advantages of EI patents
Methods for conceiving EI patents
The methods for conceiving EI patents are shown in Figure
Figure 8.1: Methods for conceiving EI patents
Opportunities to conceive EI patents
Most applicants have applied for patents for AI inventions
before training and it ends there. However, AI will begin to
resolve problems after training.
After training, it is necessary to provide opportunities to
conceive EI patents. They include opportunities after learning
more data than the previous training since AI causes products
to behave in a more effective manner after that.
We propose setting up these opportunities after training
Comparison before and after training
At these opportunities, what you should do is observe the
We propose comparing the behaviour of products before and
after training (see Figure 8.2). Hints for EI patents are
hidden in the differences in behaviour before and after. That
is why EI patents can often be better conceived by product
experts than AI specialists.
Figure 8.2: Comparison before and after learning
The methods in Figure 8.1 are effective in the case of
reinforcement learning. The methods for conceiving EI patents
depend on the type of machine learning algorithms. The types
are "unsupervised learning", "supervised learning", and
"reinforcement learning" (see Figure 9). In the cases of
"unsupervised learning" and "supervised learning", it is
necessary to analyse details of the relationship between AI
inputs and outputs.
Figure 9: Types of algorithms and methods for conceiving EI
Summary of suggestions
At the training step and runtime step in AI systems, we
introduced IF patents and claims of a production method of
trained classifiers. Furthermore, we added a solution step to
clarify EI patents. EI patents are typically related to the
behaviour of products. EI patents are very important because
the behaviour directly solves problems. Even non-AI systems
find it hard to avoid EI patent infringement. It is much easier
to discover patent infringement via behaviour than from the
configuration of AI systems.
Artificial intelligence extends human intelligence, while
human intelligence extends artificial intelligence. That is
true for intellectual property. Since the patents we proposed
relate to the human creation of technical ideas extended by AI,
we named them extended intellectual patents.
We believe that you should add our EI patents to your patent
specialises in information and communication technology
and physics. He has expertise in standard essential
patents. Nishizawa has experience in the design and
development of computer systems for solution
specialises in electronics and software and was
registered in 2015. Takahashi has experience in SoC
design and verification for Renesas Electronics
specialises in electronic and electrical engineering,
software, telecommunications, mechanical control, plant
control and business models. Furuichi was registered in
specialises in mechanical engineering and was registered
in 2004. Kajii has experience in domestic and
international patent applications and prosecution mainly
in relation to structures and controlling methods of