On December 6, 2024, China’s National Intellectual Property Administration (CNIPA) released the Draft Guidelines for Patent Applications for Artificial Intelligence-Related Inventions (人工智能相关发明专利申请指引(征求意见稿)). The Guidelines detail the type of Artificial Intelligence (AI) related patent applications; identification of inventors; subject matter eligibility; disclosure requirements; inventiveness examination; and ethical issues. Excerpts follow. The full text is available here (Chinese only). Comments are due before December 13, 2024.
I. Types of AI-related Patent Applications
The Guidelines state that there are 4 types of AI-related patent applications:
- AI algorithms of models themselves;
- applications of AI algorithms or models in different domains;
- inventions made with the assistance of AI; and
- AI-generated inventions.
II. Inventorship
The Guidelines state,
Section 4.1.2 of Chapter 1 of Part I of the Guidelines clearly states that “the inventor should be an individual, and the name of the organization or collective, as well as the name of the artificial intelligence, should not be included in the request”. The inventor whose name appears in the patent document must be a natural person, and the AI system must be named as a natural person. and other non-natural persons may not be inventors. When there are multiple inventors, each inventor must be a natural person. The property right to receive income and the personal right to sign the name enjoyed by the inventor are both civil rights, and only civil subjects in accordance with the provisions of the civil law can act as the right holders of the inventor’s relevant civil rights, and the AI system currently cannot enjoy civil rights as a civil subject, and therefore cannot act as an inventor…
For inventions made with the assistance of artificial intelligence, a natural person who has made a creative contribution to the substantive features of the invention may be named as the inventor of the patent application. For inventions generated by artificial intelligence, it is not possible to grant the status of inventor to artificial intelligence in the current legal context of China.
III. Subject Matter Eligibility
Artificial intelligence algorithms or models are developed based on mathematical theories. If the claims of a patent application involving artificial intelligence algorithms or models only involve abstract mathematical theories or mathematical algorithms and do not contain any technical features, they are rules and methods of intellectual activities and cannot be granted patent rights. For example, a method for establishing a general neural network model based on an abstract algorithm and without any technical features, or a method for training a general neural network using an optimized loss function to accelerate training convergence without any technical features, are both considered to be an abstract mathematical algorithm and are rules and methods of intellectual activities.
Scenario 1: AI algorithms or models process data with precise technical meaning in a technical domain
A solution defined by a claim is a technical solution if the claim is written in such a way as to reflect that the object processed by the artificial intelligence algorithm or model is data with a precise technical meaning in the technical field, such that, based on the understanding of a person skilled in the art, it is possible to know that the execution of the algorithm or model directly embodies the process of utilizing the laws of nature in order to solve a technical problem and that a technical effect is obtained. For example, a method for recognizing and classifying images using a neural network model. Image data belongs to the technical field with precise technical meaning, and if a person skilled in the art can know that the various steps of processing image features in the solution are closely related to the technical problem to be solved of recognizing and classifying an object, and a corresponding technical effect is obtained, then the solution belongs to the technical solution.
Scenario 2: There is a specific technical connection between the artificial intelligence algorithm or model and the internal structure of the computer system
If the drafting of the claim can reflect the specific technical connection between the artificial intelligence algorithm or model and the internal structure of the computer system, thereby solving the technical problem of how to improve the hardware computing efficiency or execution effect, including reducing the amount of data storage, reducing the amount of data transmission, improving the hardware processing speed, etc., and can obtain the technical effect of improving the internal performance of the computer system in accordance with the laws of nature, then the solution defined in the claim belongs to the technical solution.
This specific technical connection reflects the mutual adaptation and cooperation between the algorithm features and the related features of the internal structure of the computer system at the technical implementation level, such as adjusting the system architecture or related parameters of the computer system to support the operation of a specific algorithm or model, making adaptive improvements to the algorithm or model for a specific internal structure or parameters of the computer system, or a combination of the above two.
For example, a neural network model compression method for a memristor accelerator includes: step 1, adjusting the pruning granularity according to the actual array size of the memristor during network pruning through an array-aware regularized incremental pruning algorithm to obtain a regularized sparse model adapted to the memristor array; step 2, reducing the ADC accuracy requirement and the number of low-resistance devices in the memristor array through a power-of-two quantization algorithm to reduce the overall system power consumption.
In this example, in order to solve the problem of excessive hardware resource consumption and excessive power consumption of the ADC unit and the computing array when the original model is mapped to the memristor accelerator, the scheme uses a pruning algorithm and a quantization algorithm to adjust the pruning granularity according to the actual array size of the memristor to reduce the number of low-resistance devices in the memristor array. The above means are algorithm improvements made to improve the performance of the memristor accelerator. They are constrained by hardware condition parameters, reflecting that the algorithm characteristics have a specific technical association with the internal structure of the computer system, and using technical means that conform to the laws of nature to solve the technical problems of excessive hardware consumption and high power consumption of the memristor accelerator, and obtain the technical effect of improving the internal performance of the computer system that conforms to the laws of nature. Therefore, this solution belongs to the technical solution.
Specific technical association does not mean that the hardware structure of the computer system must be changed. For the solution of improving the artificial intelligence algorithm, even if the hardware structure of the computer system itself has not changed, the solution can achieve the technical effect of improving the internal performance of the computer system as a whole by optimizing the system resource configuration. In this case, it can be considered that the artificial intelligence algorithm characteristics have a specific technical association with the internal structure of the computer system, which can improve the execution effect of the hardware.
For example, a training method for a deep neural network model includes: when the size of the training data changes, for the changed training data, respectively calculating the training time of the changed training data in the preset candidate training scheme; selecting the training scheme with the shortest training time from the preset candidate training schemes as the best training scheme for the changed training data, the candidate training schemes include a single processor training scheme and a multi-processor training scheme based on data parallelism; and performing model training on the changed training data in the best training scheme.
This solution is to solve the problem of slow training speed of deep neural network models. For training data of different sizes, a single processor training solution or a multi-processor training solution with different processing efficiencies is selected. The model training method has a specific technical connection with the internal structure of the computer system, which improves the execution effect of the hardware during the training process, thereby obtaining the technical effect of improving the internal performance of the computer system in accordance with the laws of nature, thus constituting a technical solution.
However, if a claim only uses a computer system as a carrier for implementing the operation of an artificial intelligence algorithm or model, and does not reflect the specific technical connection between the algorithm characteristics and the internal structure of the computer system, it does not fall within the scope of Scenario 2.
For example, a computer system for training a neural network includes a memory and a processor, wherein the memory stores instructions and the processor reads instructions to train the neural network using an optimized loss function.
In this solution, the memory and processor in the computer system are only conventional carriers for algorithm storage and execution. There is no specific technical connection between the algorithm characteristics involved in training the neural network using an optimized loss function and the memory and processor contained in the computer system. This solution solves the problem of optimizing neural network training, which is not a technical problem. The effect obtained is only to improve the model training efficiency, which is not a technical effect of improving the internal performance of the computer system, and therefore does not constitute a technical solution.Scenario 3: Mining the intrinsic correlations in big data in specific application fields based on artificial intelligence algorithms that conform to natural laws
When artificial intelligence algorithms or models are applied in various fields, data analysis, evaluation, prediction or recommendation can be performed. For such applications, if the claim reflects that the big data in a specific application field is processed, and the intrinsic correlations between data that conform to natural laws are mined using artificial intelligence algorithms such as neural networks, the technical problem of how to improve the reliability or accuracy of big data analysis in specific application fields is solved, and the corresponding technical effects are obtained, then the scheme of the claim constitutes a technical scheme.
The means of using artificial intelligence algorithms or models to mine data and train artificial intelligence models that can obtain output results based on input data cannot directly constitute technical means. Only when the intrinsic correlations between data mined based on artificial intelligence algorithms or models conform to natural laws, can the relevant means as a whole constitute technical means that utilize natural laws. Therefore, it is necessary to clarify in the scheme recorded in the claim which indicators, parameters, etc. are used to reflect the characteristics of the analyzed object in order to obtain the analysis results, and whether the intrinsic correlations between these indicators, parameters, etc. (model input) mined using artificial intelligence algorithms or models and the result data (model output) conform to natural laws.
For example, a food safety risk prediction method obtains and analyzes historical food safety risk events, obtains various head entity data and tail entity data representing food raw materials, edible items, and food sampling poisons, and their corresponding timestamp data, and constructs corresponding quadruple data according to each head entity data and its corresponding tail entity data, and its corresponding entity relationship carrying timestamp data representing the content level, risk or intervention of various types of hazards, and obtains the corresponding knowledge graph; uses the knowledge graph to train a preset neural network to obtain a food safety knowledge graph model; and predicts the food safety risk at the time of prediction based on the food safety knowledge graph model.
The background technology of the scheme specification records that the prior art uses static knowledge graphs to predict food safety risks, which cannot reflect the fact that food data in actual situations changes over time, and ignores the influence between data. Those skilled in the art know that food raw materials, edible items, or food sampling poisons will gradually change over time. For example, the longer the food is stored, the more microorganisms are contained in the food, and the content of food sampling poisons will increase accordingly. When the food contains a variety of raw materials that can undergo chemical reactions, the chemical reaction may also cause food safety risks at some point in the future over time. This solution is based on the inherent characteristics of food that change over time to predict food safety risks, so that timestamps are added when constructing the knowledge graph, and the preset neural network is trained based on the entity data related to food safety risks at each moment to predict the food safety risks at the time to be predicted. It uses technical means that follow the laws of nature to solve the technical problem of inaccurate prediction of food safety risks at future time points, and can obtain corresponding technical effects, so it constitutes a technical solution. If the intrinsic correlation between the indicator parameters mined by artificial intelligence algorithms or models and the prediction results is only subject to economic laws or social laws, it is a case of not following natural laws. For example, a method for estimating a regional economic prosperity index using a neural network uses a neural network to mine the intrinsic correlation between economic data and electricity consumption data and the economic prosperity index, and predicts the regional economic prosperity index based on the intrinsic correlation. Since the intrinsic correlation between economic data and electricity consumption data and the economic prosperity index is subject to economic laws and not natural laws, this solution does not use technical means and does not constitute a technical solution.
IV. Adequate Disclosure
1. Determine the contents that should be recorded in the specification according to the type of invention contribution
The specification should clearly record the technical solution of the invention, describe in detail the specific implementation method of the invention, and fully disclose the technical content that is essential for understanding and realizing the invention, to the extent that the technical personnel in the relevant technical field can realize the invention.
Artificial intelligence algorithms or models have “black box” characteristics, and sufficient information is required to achieve the purpose of full disclosure. The technical content that is essential to realize the invention is different for different invention contributions.
The specification should fully describe the part that contributes to the prior art. For the technical means that embody the concept of the patent invention, the specification should clearly and completely describe it, subject to the ability of the technical personnel in the relevant technical field to realize it.
The specification should clearly and objectively state the beneficial effects of the application compared with the prior art. If necessary, corresponding evidence can be provided to prove its invention contribution.
2. Writing of application documents involving different types of invention contributions
The following recommended practices are given as examples:
For applications whose invention contributions lie in AI model training, it is generally necessary to clearly record the algorithms involved in the necessary model training process and the specific steps of the algorithms, the specific process of the training method, etc. in the specification, based on the problems to be solved or the effects to be achieved by the solution.
For applications whose invention contributions lie in AI model construction, it is generally necessary to record the necessary module structure, hierarchical structure or connection relationship in the specification, and accurately and objectively describe the functions and effects of the model, based on the problems to be solved or the effects to be achieved by the solution. If necessary, the effects that can be achieved after the improvement can be demonstrated through experimental data, analysis and demonstration, etc.
For applications whose invention contributions lie in the application of AI in specific fields, it is generally necessary to clarify in the specification how the model is combined with the specific application scenario, how the input/output data is set, etc., based on the problems to be solved or the effects to be achieved by the solution. If necessary, the description should also clarify the correlation between the input data and the output data, so that the technicians in the relevant technical field can judge that there is a correlation between the two.
For the review opinion that the description is not sufficiently disclosed, the reasons and basis for the technicians in the relevant technical field to implement the relevant solutions need to be explained when stating the opinion. It should be noted that whether the description is sufficiently disclosed shall be based on the contents recorded in the original description and claims.
V. Inventiveness Examination
In order to make the artificial intelligence algorithmic features be included as part of the technical means in the judgment of inventiveness, the claims should reflect that the artificial intelligence algorithms or models are applied to specific functions or application fields, and solve specific technical problems, so as to make it clear that the algorithmic features and the technical features functionally support each other and exist in an interactive relationship, so as to make the algorithmic features become an integral part of the technical means.
…
If the artificial intelligence algorithm or model recorded in the application scheme belongs to the prior art, and the improvement of the scheme lies in applying it from the existing technical field to the technical field of this application, then the inventiveness consideration should comprehensively consider the distance of the technical field to which the algorithm or model is applied, whether there is a corresponding technical inspiration, the difficulty of applying to different field scenarios, whether it is necessary to overcome technical difficulties, and whether it brings unexpected technical effects.
Further, if the algorithm or model is applied to different field scenarios, and the training method, parameters, configuration and other elements of the algorithm or model are not adjusted by overcoming technical difficulties, and unexpected technical effects are not obtained, then the scheme cannot be inventive.
For example, an application involves a method for counting the number of ships. Based on the ship image data, a real-time detection data model is trained through deep learning, and the number of detected ships is summed to solve the technical problem of real-time feedback of the number of ships in the current sea area. The closest prior art discloses a method for counting the number of fruits on a tree, and discloses the deep learning model training and quantity counting steps of the application. The difference lies in the different recognition objects, which belong to different application scenarios. Although ships and fruits differ in appearance, volume, and the environment in which they exist, for those skilled in the art, the means used by both are to perform object recognition and model training on the obtained image information, and then complete the quantity statistics. When recognizing the image, the position and boundary of the recognized object are also considered. If the recognition and training of ships in the image and the recognition and training of fruits in the image do not change the processing methods in deep learning, model training process, and image recognition, and the technical effects that can be obtained are to make the statistical results more accurate, then the difference in training data only means that the data has different meanings, and the difference in data meanings does not constrain, influence, or limit the improvement or implementation of the algorithm, and the difference in application scenarios does not impose different constraints, influences, or limitations on the design of the algorithm model. Therefore, the effect of applying the fruit statistics method of the prior art to the ship statistics of this scheme is predictable based on the prior art, and no unexpected technical effects are produced, and the scheme does not have inventiveness.…
If there is a specific technical relationship between the artificial intelligence algorithm or model and the internal structure of the computer system, and the internal performance of the computer system is improved, the algorithm features and technical features in the scheme will be considered as a whole when judging the inventiveness.
The situation of improving the internal performance of the computer system includes: supporting or optimizing the operation of a specific algorithm or model by adjusting the architecture of the hardware system, optimizing the scheduling of hardware resources in the computer system by executing the algorithm or model, etc. In this case, the algorithm features and technical features in the scheme will be considered as a whole. If the prior art does not provide technical inspiration, the scheme is inventive.
For example, an application involves a method for adjusting a convolutional neural network, which reduces resource usage by fixing the neural network, so that a neural network model with low-bit fixed-point quantization can run on a low-bit width FPGA platform, and can achieve calculation accuracy comparable to that of a floating-point network under low bit width. The closest prior art discloses a fixed-point training method based on dynamic fixed-point parameters for a convolutional neural network. In the training process of the convolutional neural network, this method uses a fixed-point method for forward calculation, and within several training cycles, the network accuracy reaches the level of floating-point calculation. The difference between this solution and the closest prior art is that after the convolutional neural network is trained using high-bit fixed-point quantization, the convolutional neural network is fine-tuned through the low bit width of the FPGA. Based on this distinguishing feature, this application solves the technical problem of using multi-level, large-scale convolutional neural networks in small FPGA embedded systems to overcome limited resources, reduce the resource usage of convolutional neural networks trained on the FPGA platform, and obtain the technical effect of improving the internal performance of the computer system. Considering the algorithm features and the technical features such as the low bit width of the FPGA as a whole, there is no technical inspiration in the prior art, and this solution is inventive.If the artificial intelligence algorithm features and technical features in the scheme improve the user experience together, the algorithm features and technical features will be considered as a whole when judging the inventiveness. If the prior art does not provide technical inspiration, the scheme is inventive.
For example, an application involves a method for implementing online customer service to solve the technical problem that users in the existing e-commerce platform tend to handle complaints and consultations through manual customer service, resulting in the unreasonable use of chatbot customer service and manual customer service resources, and the pressure on manual customer service. The main solutions adopted include: using long short-term memory networks to analyze the context of user requests, and combining genetic algorithms to optimize the dynamic allocation of manual and chatbot customer service. When it is detected that the manual customer service is overloaded, the system uses long short-term memory networks to predict and automatically direct suitable requests to chatbot customer service to reduce the processing pressure on manual customer service. The closest prior art discloses a method for implementing chat with online customer service, which specifically discloses that users can freely choose and switch between three ways to communicate with customer service: only chatbot customer service, chatbot customer service priority, and manual customer service priority. In the “manual customer service priority” mode, when the upper limit has been reached or there is a queue, the chatbot customer service will communicate with the user. The closest existing technology mainly switches between manual or chatbot customer service based on user selection, and the basis for judging whether the manual customer service is busy is whether the limit has been reached or whether there is a queue. This is different from the automatic switching performed by the present application after weighing the artificial intelligence algorithm. This solution analyzes the access load through an artificial intelligence algorithm and automatically switches to robot customer service. It can solve the technical problem of more reasonably allocating user service requests between chatbot customer service and manual customer service, save user waiting time, and improve user experience. Therefore, the solution is inventive.
VI. Ethics
The continuous development of Artificial Intelligence brings more development opportunities to various industries, and also brings ethical issues such as algorithmic ethics, data security, and data compliance. Patent applications on AI-related content should comply with the provisions of Article 5 of China’s Patent Law.
If the application of AI algorithms or models in different fields is involved, the applicant should pay attention to whether the algorithms or models are in violation of relevant laws and regulations, social morality, or detrimental to the public interest when they are applied to specific scenarios. Where artificial intelligence is involved in the acquisition and utilization of data, attention needs to be paid to whether the source of the data in question, application scenarios, safety management, use norms and other aspects follow relevant laws and regulations. In addition to the content of the data itself, the specific means of data collection, storage, processing, etc. also need to comply with the requirements of relevant laws and regulations, and must not violate social morality or jeopardize the public interest.
