Case Studies of AI-related Inventions in Taiwan’s Patent Examination Guidelines

The new “Examination Guidelines on Computer Software-Related Inventions” recently came into effect, addressing many issues within the established examination practice regarding Computer Software-related (CSR) Inventions (see full analysis here). One of the issues addressed by the revision this year was the lack of examples to illustrate specific patentability requirements. The new guidelines provided numerous examples (along with detailed explanations) that help readers understand specific requirements and how they could be fulfilled. Readers deeply concerned about the patentability of their inventions may be pleasantly surprised by the wide variety of inventions included in these examples. In particular, examples related to Artificial Intelligence-related (AI-related) inventions, perhaps one of the most contested subject matters regarding patentability, have been thoroughly discussed in these examples. We will take a closer look at some of these examples in this post.

Definition of Invention

The new guidelines provided a clearer process of examination that is consistent with the general definition of inventions set forth in the Patent Act (Article 21). The examination process can be seen in Fig. 1 (for more information about the examination procedure, please refer to our post here).

Fig. 1 Decision Flowchart of Eligibility of Software-Related Inventions

To help decipher many of these requirements, the guidelines provide examples of AI-related inventions that satisfy or fail the test.

Case study 2-12 Neural Network for Analyzing and Rating Accommodation Reputation

The claim of the invention is as follows:
A neural network system for analyzing accommodation reputation, the system causing a computer to, based on text-based data regarding accommodation reputation, produce multiple quantitative metrics about the accommodation reputation, the system comprising:
    a first neural network and a second neural network, wherein input of the second neural network is output of the first neural network;
    the first neural network being a network that is comprised of an input layer and intermediate layers of a feature extraction neural network, wherein at least one intermediate layer of the feature extraction network has fewer neurons than the input layer, the number of neurons in the input layer is equal to that of an output layer of the feature extraction neural network, and weights of the feature extraction neural network are trained in such a way that input value to the input layer gradually becomes equal to output value of the output layer;

    weights of the second neural network being trained on the premise that the weights of the first neural network remain unchanged;
    the system causing the computer to perform calculation based on the trained weights of the first and second neural networks in response to frequency of appearance of specific phrases in the text-based data regarding accommodation reputation that is inputted to the input layer of the first neural network, and to output the multiple quantitative metrics about the accommodation reputation from an output layer of the second neutral network.

The claim is compatible with the definition of invention.

The claim describes a technical means that relies on the collaboration of computer software and hardware resources, with the purpose of information processing (accurate analysis of accommodation reputation), to achieve specific information processing and computation (the training of the weights of the first and second neural networks and, utilizing the resulting weights to perform calculation in response to the frequency of appearance of keywords in the text-based data inputted into the first neural network and to output multiple quantitative metrics from the output layer of the second neural network) to construct a specific information processing system. This claim is compatible with the definition of invention.

Enablement Requirement

The enablement requirement requires that a claim and its specification enable a person ordinarily skilled in the art to understand a claimed invention and make use of the invention. This requires that the specification be clear and have sufficient disclosure. The guidelines provided some guidance as to how AI-related inventions should address this requirement.

Case study 1-2 Deep Neural Network-based Matching System for Real Estate Agents

The claim of the invention is as follows:
A real estate agent matching system utilizing deep neural network, comprising:
    a database that stores client data, real estate data and agent data; the client, real estate, and agent data of completed cases being labeled and used as training data; and
    a deep neural network model that connects to the database, the deep neural network model being trained by using the client and real estate data of the completed cases as inputs and the real estate agent data of the completed cases as corresponding outputs;

    wherein the trained deep neural network model takes any client and real estate data as input and outputs agent data that corresponds to the inputted client and real estate data.

The gist of the specification/description:
Whether a real estate transaction can successfully take place strongly depends on the agent, especially the agent’s professional knowledge of specific types of real estate and his/her service. If an optimal matching among clients, real estate, and agents can be achieved, it can spur the growth of the real estate market.

The present invention provides a deep neural network (DNN)-based matching system for real estate agents that includes a database that stores client, real estate, and agent information. The training data of the model consists of completed transactions with the respective clients, real estate, and agents. The resulting model can predict the optimal matching for any client, real estate, and agents to maximize the likelihood of a successful transaction.

The invention does not fulfil the enablement requirement.

While the specification disclosed the client, real estate, and agent data as training data, it did not mention the correlation between the three types of data, or the specific technical means of how optimal matching could be achieved from the machine learning model (e.g., data preprocessing, the size of the training data, the type of neural network, loss function, and validation). Even though, considering the general knowledge at the time of filing the application, one could reasonably know that client/agent data includes photo, name, age, and address, and real estate data includes size and year built, one could still not find a relationship among the aforementioned data. Thus, it is hard to determine whether the trained model based on these data could achieve optimal matching. Therefore, the claimed invention does not fulfill the enablement requirement.

Clarity and Sufficiency of Disclosure

Related to the enablement requirement, the guidelines also specified how specifications and claims should be drafted for CSR inventions. The guidelines further provided examples of claims that are unclear (“indefinite”) and therefore could face invalidation challenges. Examples of AI-related inventions regarding this requirement shed light on how claims of AI-related inventions should be drafted. Here is an example of a claim that is considered “indefinite” under the scenario “ambiguity resulting from the claim language” (see all of the scenarios of “indefiniteness” here):

Example 1 of Ambiguity Resulting from the Claim Language

The claim of the invention is as follows:
A translation machine comprising: a high speed vocabulary processing device; and a grammar processing device; wherein the two devices are capable of processing in parallel.

Neither the specification nor the claims specify the meaning of “high speed.” Even with the consideration of general knowledge at the time of filing the application, “high speed” is still unclear in terms of its reference of comparison/relativity or degree, and therefore the scope of this claim is unclear and cannot be ascertained. The claim is therefore indefinite.

Inventive Step (Non-Obviousness)

The inventive step requirement is perhaps the most important requirement in determining the patentability of an invention. Due to the many unique properties of CSR patents, the guidelines provided many factors used by examiners to determine whether a CSR invention has inventive step. The following is an example of how inventive step is evaluated regarding an AI-related invention.

Case study 3-5 Traffic Flow Estimation System

The claim of the invention is as follows:
A system for estimating a number of vehicles passing through an ETC toll gate at a particular road segment, comprising:
    a data reading unit capable of reading from an ETC control station historical data of a number of vehicles that passed through the ETC gate at the particular road segment;
    a neural network unit built by a processor and trained in advance using the historical data; the neural network unit having an input layer and an output layer; the input layer receiving the historical data of the number of vehicles that passed through the ETC gate at the particular road segment prior to a specified point in time as input; the output layer outputting a prediction of the number of vehicles that will pass through the ETC gate at the particular road segment at that specified point in time;
    a comparison unit for determining a difference between the prediction outputted by the neural network unit and an actual number of vehicles that passed through the ETC gate at that specified point in time; if the difference is greater than a predetermined threshold, the comparison unit determining the number of vehicles at the particular road segment as abnormal; and

    a map labeling unit that automatically labels or provides warning indication on an electronic map for the particular road segment that is determined as abnormal.

The primary cited reference is as follows:
A recursive model-based traffic flow estimation system that could receive historical data of the number of vehicles from toll booths and predict the number of vehicles that pass by in real-time.

Other cited reference:
A system that utilizes neural network to predict pedestrian flow in a train station; the system relies on historical data of pedestrian flow as training data for a neural network model and uses the trained model to predict pedestrian flow at a specified point in time. The system incorporates an electronic map that signals warning real-time when the pedestrian flow is abnormal to help adjust the capacity of the station and the frequency of trains coming in and out of the station.

The clamed invention does not have inventive step.

The difference between the claimed invention and the primary cited reference is that the cited reference did not utilize a neural network in predicting traffic flow and did not disclose incorporating the resulting prediction in an electronic map to provide warning of abnormality. However, this difference is addressed in the other cited reference. As both the primary cited reference and the other cited reference belong in Computer Software-related technical area, and both utilize mathematical models on issues of traffic control, there’s commonality with respect to the technical area, problem to be solved, and function or effect. A person of ordinary skill in the art would therefore be motivated to combine the primary cited reference with the other cited reference. Moreover, it is common to utilize neural networks to increase accuracy of prediction; a person of ordinary skill in the art would therefore be motivated to use neural networks alongside, or in place of, the recursive model disclosed by the primary cited reference, in addition to combining the prediction result with an electronic map to provide warning, thereby easily resulting in the claimed invention. The claimed invention therefore does not have inventive step.

The examples from the revised guidelines provide readers with additional guidance on how patent applications for CSR inventions should be drafted. The diverse set of technologies used as examples (even within AI, ranging from autonomous vehicles to natural language processing) also illustrates the TIPO’s openness towards new and emerging AI technology.