Author: Wang Yong Senior Patent Agent
Traditional neural networks generally refer to the network formed by the mutual connection of nerve cells (i.e., neurons) in the biological body through the protrusions on the cell body, which is used to generate the consciousness of the biological body and helps the biological body to think and act. Artificial neural networks (Artificial Neural Network, hereinafter referred to as neural networks) are mathematical models that simulate the connection network of neurons in the biological body and can mimic the intelligent activities of the brain of the biological body to perform mathematical calculations, thereby solving various computational tasks. It can be seen that neural networks are a kind of mathematical algorithm that naturally has the attribute of "intelligent" and is universal. This also leads to the fact that all or part of the content of the patent application related to neural networks is easily regarded as the rules and methods involving intellectual activities, and further is considered not to belong to the technical scheme and is excluded from the patent authorization object.
In view of the patent applications for neural networks or other new business models and new fields involving artificial intelligence, the State Intellectual Property Office issued the Announcement on Amending the Patent Examination Guidelines on December 31, 2019 (No. 343). The amended Examination Guidelines for Patents stipulate the particularity of examination for invention patent applications involving artificial intelligence, “Internet+”, big data, and blockchain, and clarify the examination rules for related patent applications to a certain extent through several examination examples. Specifically, the three elements of the technical means, technical problems, and technical effects used in the determination of the subject matter are detailed explained. However, in patent practice, due to the diversity and complexity of the scheme, the related application content of neural networks is usually difficult to accurately delimit between the abstract mathematical algorithm and the specific technical scheme. When facing the issue of whether a patent application belongs to the authorized subject matter stipulated in the Patent Law, patent applicants/patent agents and patent examiners often come to different judgment results due to cognitive differences.
When the patent examination officer issues a rejection decision that the claimed invention of a neural network-related patent application does not belong to the patentable subject matter in accordance with the relevant provisions of Article 2, Paragraph 2, or Article 25, Paragraph 1, of the Patent Law, the patent applicant may request a review by the Patent Reexamination Board in accordance with Article 41 of the Patent Law. The Patent Reexamination Board may issue a review decision to uphold or overturn the original rejection decision after the review. The patent reexamination procedure can correct the cognitive errors of the relevant subjects in the patent examination procedure in some cases, and the results of the patent reexamination can also reflect the applicable conditions stipulated in the relevant provisions to a certain extent more objectively. Therefore, this paper makes a simple discussion on the issue of the object of neural networks based on a number of review cases, in order to provide a certain reference for the drafting of neural network-related patent applications and the reply to examination opinions involving the issue of the object.
1. Three perspectives on the ontological problem of neural networks
The protection subjects related to neural networks mainly involve the model structure of neural networks, the model training/Optimization method of neural networks, or the use method of neural networks in specific scenarios and other types of subjects. Various different types of neural network-related solutions all need to rely on computer equipment to achieve; for example, the model structure of the neural network is generally installed in the memory of the computer equipment, and the neural network model framework, network parameters, and input/output data can be saved in different storage areas of the memory; for example, the training method/application method of the neural network is executed by the processor of the computer equipment, and the processor of the computer equipment needs to read/write the network parameters and related input/output data from the memory when executing the relevant neural network algorithm. So, is the neural network combined with the computer equipment enough to constitute the patent protection object stipulated in our country's patent law? Perhaps we can find the answer from the past domestic and foreign patent practice.
In the patent examination and judicial practice in the United States and Europe, the objectivity judgment of neural networks is generally more relaxed than in our country. For example, Google and its subsidiary DeepMind have submitted a large number of patent applications related to neural networks in the United States, among which there are no lack of related schemes that are generally considered to be basic algorithms or public resources, such as Dropout (a method of preventing overfitting in neural network training), CNN (convolutional neural network), RNN (recurrent neural network), etc.
In fact, the United States has always been the country with the broadest scope of patentable subjects, and the US patent law stipulates that any new and useful method, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may be patented under the conditions and requirements specified in this law. In the judicial practice of the United States, the "machine or transformation test" has been proposed and used to guide the issue of patentability of methods, which refers to the fact that a method as a patentable subject needs to be combined with a specific machine or can transform one item into another different state or another thing. Based on this test method, the neural network algorithm combined with the computer equipment obviously meets the conditions of being combined with a specific machine and transforming the input data.
The European Patent Office (EPO) performs a division of the features in the claims into technical and non-technical features during the patent examination process, with the non-technical features being excluded from the evaluation of novelty and inventiveness. In determining patentability of the claims, the EPO has used the "contribution approach" (contribution approach) for evaluation, which stipulates that only the parts of the claim that contribute to the state of the art should be considered in the patentability evaluation. However, due to the violation of the principle that the claim should be evaluated as a whole, the contribution approach was soon abandoned. On this basis, the EPO established the principle that as long as the claim includes a computer, a computer network, or a computer readable medium, the claim will be attributed with a technical character, which is the implicit requirement that the European Patent Convention requires for an invention to be patentable. Based on this evaluation principle, when a neural network algorithm is combined with a computer device, it can be preliminarily determined that it meets the patentability requirements for technical character.
In the practice of patent in our country, when facing the issue of whether a patent application related to neural networks belongs to the authorized objects specified in the Patent Law, there may be the following three kinds of cognition in general:
(1) Artificial neural networks belong to a kind of universal algorithmic tool. The relevant structure or method of artificial neural networks relies on the hardware structure of computer equipment, and its data processing process inevitably involves the technical means of data acquisition, data transmission, data storage and so on inside the computer. Therefore, artificial neural networks themselves are a kind of technical scheme that can be applied in a variety of different technical fields, and belong to the objects of patent protection.
(2) The clear delimitation of a specific technical field in a claim involving a neural network, based on the model structure or model training/optimization/application method of the neural network in that technical field, involves a technical means for processing technical data in that technical field by a computer program. Such a scheme belongs to the invention patent application specified in Chapter 9 of the Examination Guide involving a computer program, specifically involving the processing of external technical data by executing a computer program. Therefore, the neural network combined with a specific technical field constitutes a technical scheme, which belongs to the patent-protected object.
(3) In the absence of a specific technical field, optimizing the structure or algorithmic steps of a neural network involves techniques for performance optimization of a computer program in terms of computational resources and computational efficiency. Such a scheme falls under the invention of a computer program as defined in Chapter 9 of the Examination Guidelines, specifically involving the improvement of internal performance of a computer system by executing a computer program. Therefore, even without combining with a specific technical field, a scheme for enhancing computational performance based on a neural network is sufficient to constitute a technical scheme and belongs to the subject matter protected by patents.
With regard to the three cognitive positions stated above, the examination results of the specific reexamination cases are analyzed one by one below. It should be noted that the following reexamination cases only briefly quote the relevant content of the specific reexamination decisions. For the complete content of the reexamination decisions of the relevant cases, please refer to the official website of the Patent Reexamination and Invalidity Examination Department of the State Intellectual Property Office.
Second, can neural networks involving multiple technical fields be protected as a universal technical solution and become the subject matter of patent protection
Case 1: Sparse adaptive neural network, algorithm and implementation device for deep learning
Decision No. 192123 Decision Date 2019-09-23
Application number 201510944909.7; Application date 2015-12-16
The refused decision in case 1 concerned claim 1 as follows:
A sparse adaptive neural network algorithm for deep learning, implemented on a single chip, characterized by:
1) The convex optimization formulation of the objective function is represented in the negative log-likelihood form of the maximum likelihood estimation of the probability distribution of the energy function;
2) The objective function mentioned above is: , where λ represents the regularization coefficient, l represents the number of layers in the network, the minimum is 1 layer, v(l) represents the visible layer of each layer of RBM, h(l) represents the hidden layer of each layer of RBM; adding a one norm regularization term to the objective function, , wij represents the weight of the connection;
3) The optimization target of the objective function mentioned above is x ≤ wi,j ≤ y, x, y ∈ R, and the optimization result is that wij is close to the endpoints of the optimization target interval x, y;
4) The weights of the connected edges are represented by short bit-width discrete integer values;
5) Round x, y to the nearest integer, represented by [x], [y], round [x] ≤ m1, m2 ≤ [y], when wi,j ≤ m1, wi,j = [x], when wi,j ≥ m2, wi,j = [y], otherwise wi,j is represented by a conventional value indicating the connection does not exist;
6) After the gradient descent step of the conventional item, the contrastive divergence rule is applied for updating, and the parameters are updated with the gradient of Rs.
It is stated in the decision to reject that claim 1 seeks to protect a sparse adaptive neural network algorithm for deep learning, in which the scheme is to achieve a sparse representation by limiting the number of active hidden units. It is merely an improvement of the algorithm itself and is not applied to any technical field. It is not mentioned what practical technical problems the algorithm can be used to solve, nor is it limited to the physical meaning of the algorithm parameters in the technical problem, that is, it does not form a technical scheme with technical significance, nor does it reflect the technical effects that can be brought about by the use of the algorithm to solve the technical problem; it belongs to the rules and methods of intellectual activities.
In the reexamination procedure, the applicant for reexamination stated: "The neural network algorithm for deep learning" itself is a technical solution, because the neural network model is applicable to integrated circuits, and the specific fields of use are extensive, such as in the application of machine vision, scene perception, feature extraction, big data processing, etc., considering the needs of the scope of protection, although the parameters involved in claim 1 are not specifically given physical meanings, it is routine technology for a person skilled in the art to assign physical meanings to the parameters in the actual application process.
In response to the statement of the applicant for reconsideration, the examination team believes that: artificial neural networks are a kind of algorithmic mathematical model that imitates the behavior characteristics of biological neural networks and carries out distributed parallel information processing. It is only a mathematical model itself and does not constitute a technical solution. Only when it is applied to specific application fields, and the data in the specific application field is processed to solve the technical problems such as classification and recognition in the specific field, the whole scheme belongs to a technical solution. The applicant for reconsideration listed multiple application fields for this application, which actually shows that this application is not combined with a specific application field, but only an abstract mathematical model.
The panel affirms the decision of the State Intellectual Property Office to reject this application.
Case 2: Hierarchical neural network device, discriminator learning method, and discrimination method
Decision No. 195731 Decision Date 2019-10-15
Application number 201480073042.6; Application date 2014-02-10
The rejected claim 1 of example 2 is as follows:
An apparatus for discrimination using a hierarchical neural network for control, prediction, or diagnosis, the hierarchical neural network being:
Weight storage part, which stores the weights between the nodes in a hierarchical neural network;
Weight learning unit, which learns the weight between a plurality of nodes in the hierarchical neural network, the hierarchical neural network is formed by sparse coupling between the nodes based on a parity check matrix of an error correcting code, and comprises an input layer, a middle layer, and an output layer each having the nodes; and
Discrimination processing section, which uses a hierarchical neural network that updates the weight values of the nodes coupled after learning the weight values by the weight learning section to perform discrimination, to solve a classification problem or a regression problem.
It was stated in the decision to reject: claim 1 does not explicitly limit the specific technical problem that this method addresses, does not combine with a specific technical field, and the problem is solved by a method algorithm that is conducted through human thinking activities, which belongs to a mathematical problem rather than a technical problem. The actual means adopted are custom-made correction values and weight value configuration, which belong to human regulations and are not technical means; the effects brought by its scheme are also algorithm improvement bringing algorithm/mathematical effects, rather than technical effects, therefore, the solution protected by claim 1 does not constitute a technical solution, and does not comply with the provisions of Article 2, Paragraph 2 of the Patent Law.
In the reexamination procedure, the reexamination applicant added the features related to data processing, such as "the device is used for information processing related to control, prediction, or diagnosis" and "learning data storage part," "training data storage part," "discrimination processing part," etc., in claim 1, and stated the following opinions: the amended claim 1 clarified the technical field in which the device is applied, that is, "the device is used for information processing related to control, prediction, or diagnosis"; and the features such as the storage, learning, and transmission of weights and data all belong to technical means.
In view of the amendments and statements of the applicant for the reexamination, the examination division considers that: with regard to "the device is used for information processing related to control, prediction or diagnosis", it is a functional attribute of the algorithm mathematical model, not a specific technical field in the sense of the patent law. As for the specific structure of the neural network for the storage, learning and transmission of weights and data and other features, it is a functional description of the elements of the algorithm mathematical model, which belongs to the attribute itself of the algorithm mathematical model and does not belong to the technical means in the sense of the patent law.
The panel affirms the decision of the State Intellectual Property Office to reject this application.
Based on the re-examination results of Case 1 and Case 2, the following comparative analysis is conducted:
In the above two cases, the appellants limited the claimed schemes to be protected in multiple technical fields by stating opinions or amending the claims, which made the relevant neural networks a general algorithm that can be applied across multiple technical fields, unable to reflect the "specialty" of neural networks in specific technical fields. The examination division issued a review decision to uphold the rejection decision for both review requests. Specifically:
In case 1, the applicant for reconsideration stated in the statement of opinion in the reconsideration procedure that the scheme to be protected can be applied to machine vision, scene perception, feature extraction, big data processing and other technical fields, while the examination board considered that the multiple application fields listed by the applicant for reconsideration for the application of this application, on the contrary, show that this application is not combined with a specific application field, but only an abstract mathematical model, and the examination board thus drew the conclusion of maintaining the decision to reject.
In case 2, the applicant for reconsideration limited the scope of the claims to be protected by amending the claims to specifically limit the scheme to the application field of information processing related to control, prediction, or diagnosis, and such field limitation was considered a functional attribute of the algorithm data model because it was too broad and abstract, and the board of appeal thus came to the conclusion of maintaining the decision to reject.
Neural networks rely on integrated circuits or computer equipment to implement algorithms for data storage, transmission, and processing. Due to their universality, neural networks can be applied in a variety of different technical fields. In some specific contexts, the relevant schemes of neural networks can indeed be included in the broad sense of "technical schemes," but this broad sense of "technical schemes" does not strictly equate to technical schemes in the sense of patent law. If, in the statement of opinion or in the claims, a wide range of different technical fields are extensively limited, it would instead lead the examiner-panel to conclude that the overall scheme belongs to a general mathematical algorithm. Therefore, without other "patentable" basis support, limiting the neural network, which is applied in a variety of different technical fields as a general technical scheme, is not sufficient to constitute a patentable subject matter in accordance with the provisions of patent law.
Three, whether a neural network involving a specified technical field can become an object of patent protection based on the purpose of processing external technical data by executing a computer program
Case 3: The method of particle swarm optimization LVQ neural network and perturbation, harmonic detection method
Decision No. 209524 Decision Date 2020-04-15
Application number 201510310098.5; Application date 2015-06-08
The refused decision in case 3 concerned the following content of claim 1:
An online detection method for a type of distributed power source perturbation, which is characterized in that it comprises: a method of using a particle swarm optimization learning vector quantization LVQ neural network to detect the perturbation type of the input electric power quality signal based on the perturbation type information in the electric power quality signal.
wherein, a process of learning vector quantization LVQ neural network by using particle swarm optimization, comprises:
Establish a particle group, in which the components of the position of the particles in the particle group correspond one-to-one with the connection weights in the LVQ neural network;
Using a particle swarm iterative algorithm, the positions and velocities of all particles in the particle swarm are updated iteratively;
wherein each time the position and velocity of the particles are updated, the optimal fitness position of each particle is calculated, and the optimal fitness position of the particles is used to obtain the optimal fitness position of the particle group, and the optimal fitness position of the particle group is used to update the connection weight of the LVQ neural network;
When the particle swarm iterative algorithm reaches the preset number of iterations, or, the difference between the actual output and the expected output of the LVQ neural network meets the preset range, the updating of the particle position and velocity in the particle swarm is stopped;
The optimal fitness position of each particle is calculated after updating the position and velocity of the particles each time, including:
After each update of the position and velocity of the particles, calculate the current fitness value of the particles;
Check whether the current fitness value of each particle is better than the fitness value corresponding to its current optimal fitness position, if so, replace the current optimal fitness position of the particle with its current position;
The current fitness value of the所述计算 particle, including:
Use the formula to calculate the current fitness value of the particle;
where, N is the total number of training samples; yih and yia are the desired and actual outputs of the output layer corresponding to the i-th training sample, respectively.
During the substantive examination procedure, the examiner issued a rejection decision based on the lack of creativity in accordance with paragraph 3 of Article 22 of the Patent Law. In response to this rejection decision, the applicant filed a request for reconsideration.
In the reexamination procedure, the examining division proposed: Although the type of disturbance of the power quality signal is limited in the claims, it is not reflected in the claims that the type of disturbance of the power quality signal is detected based on the information of the disturbance type or that the harmonic and interharmonic disturbances of the power quality signal are detected based on the harmonic information in the input power quality signal, that is, the power quality parameters do not correlate with other features of the LVQ neural network, and therefore it still only optimizes the algorithm and does not constitute a technical scheme. This application does not comply with the provisions of Article 2, Paragraph 2 of the Patent Law, and is an object that is not authorized by the Patent Law.
The panel affirms the decision of the State Intellectual Property Office to reject this application.
Case 4: A neural network training method and device
Decision No. 206423 Decision Date 2020-03-18
Application number 201710450211.9; Application date 2017-06-15
The refused decision in case 4 is directed to claim 1 as follows:
A neural network training method, characterized in that it comprises:
Select a teacher network that performs the same function as the student network;
iteratively training the student network to obtain the target network based on the data similarity between the first output data corresponding to the same training sample data and the data similarity between the second output data to realize the transfer of the data similarity between the output data of the teacher network to the student network;
Wherein: the first output data is data output from a first specific network layer of the teacher network after the training sample data is input into the teacher network, and the second output data is data output from a second specific network layer of the student network after the training sample data is input into the student network.
It was stated in the decision to refuse: claim 1 does not involve any technical field of application, the problem to be solved is the model training itself, and the input and output of the neural network used in the model training process do not limit the specific physical parameters of a specific technical field, so the method to be protected in substance is only limited to a simple algorithm, and therefore the实质 of the claim to be protected is a rule and method of intellectual activity, which belongs to the scope of intellectual activity rules and methods referred to in Article 25, Paragraph 1, Item 2 of the Patent Law.
In the reexamination procedure, the reexamination applicant added the feature “in a real-time computer vision processing process, a processing device with low computing power acquires image data; the processing device uses a pre-set target network to perform computer vision processing on the acquired image data to obtain a computer vision processing result; where the target network is obtained by the following processing” to claim 1, and at the same time, modified “sample data” in the specification to “sample image data”.
In view of the amendments and statements of argument made by the applicant for reconsideration, the examination division considers that: the features added by the applicant for reconsideration belong to technical features, and thus the claim as a whole is not a rule or method of intellectual activity, and should not be excluded from the possibility of obtaining a patent right in accordance with Article 25 of the Patent Law. The scheme requires the processing of image data by low-compute processing equipment, and the execution of data collection and data processing such as "processing" and "obtaining" belongs to the use of technical means that follow the laws of nature to solve technical problems and obtain technical effects, in accordance with the second paragraph of Article 2 of the Patent Law.
The panel decided to set aside the decision of the State Intellectual Property Office to reject this application.
Based on the re-examination results of Case 3 and Case 4, the following comparative analysis is made:
In the above two cases, the amended claims in the appeals both include external technical data with precise technical meaning related to the neural network, and the difference lies in whether the features involving the data processing of the neural network are closely related to the external technical data, which also leads to the appeal board to issue a decision to maintain the rejection and withdraw the rejection in the two appeal cases. Specifically:
In case 3, although the applicant for reconsideration limited the "electric power quality signal", which has an exact technical meaning, in the claims, the various algorithm steps of the neural network in its scheme were not associated with the electric power quality signal, resulting in the actual processing object of its neural network (LVQ neural network) still being abstract general data. Therefore, its overall scheme does not constitute a technical scheme in the sense of the patent law, and the examination division thus came to the conclusion of maintaining the rejection decision.
In case 4, the applicant for reconsideration limited the features of obtaining and processing image data through low-computing-power processing equipment in the claims, and modified the actual processing objects of the neural network (teacher network and student network) from "sample data" to "sample image data", which reflects the close relationship between the model characteristics of the neural network and the external technical data with precise technical meaning, thus constituting a technical solution in the sense of the patent law, and the collegial group thus came to the conclusion of revoking the rejection decision.
When a technical feature related to external technical data is recorded in the claims belonging to a specified technical field (such as the power quality signal in case 3 and the image data/ sample image data in case 4), it can only be determined that the scheme as a whole does not belong to the rules and methods of intellectual activities specified in Article 25 of the Patent Law based on this technical feature, but it is also necessary to further determine whether the scheme belongs to the technical scheme specified in Article 2, Paragraph 2 of the Patent Law. Specifically, it is necessary to analyze whether the relevant features of the neural network (such as the network layer or method step of the neural network) are closely related to the external technical data, so as to determine that the actual processing object of the neural network is the external technical data with specific technical meaning or the abstract general data. On the premise that there is no other "patentability" basis to support, if a neural network-related scheme does not reflect the close relationship between the relevant features of the neural network and the external technical data, the scheme is not sufficient to constitute a patentable subject matter in accordance with the provisions of the Patent Law.
IV. Can neural networks that do not fall within a specific technical field become patentable subject matter based on the purpose of improving the internal performance of computer systems?
Case 5: Methods and devices for compressing neural networks
Decision No. 223739 Decision Date 2020-08-13
Application number 201711473963.3; Application date 2017-12-29
The rejected decision in case 5 concerns claim 1 as follows:
A method for compressing a neural network, comprising:
Obtain a trained neural network to be compressed;
Selecting at least one layer from the layers of the neural network as a layer to be compressed;
The following processing steps are performed in turn for each of the layers to be compressed, in descending order of the level number of the layer in the neural network: a pruning ratio is determined based on the total number of parameters included in the layer to be compressed, and based on the pruning ratio and the parameter value threshold, parameters are selected from the parameters included in the layer to be compressed for pruning, and the pruned neural network is trained based on the preset training samples using a machine learning method;
The neural network obtained by performing the above processing steps on the respective selected layers to be compressed is determined to be the compressed neural network and is stored.
It is pointed out in the decision of rejection: the problem to be solved by this application is how to reduce the complexity of the neural network algorithm, which is not a technical problem. In order to solve this problem, the means adopted in the claim scheme is to select different layers of the neural network and make corresponding appropriate pruning, and then train the model to achieve the effect of reducing the complexity of the algorithm while keeping the accuracy as similar as possible to the original neural network algorithm. The essence is that the non-technical means adopted has achieved the corresponding non-technical effect. It just uses non-technical means to optimize the algorithm. Therefore, the scheme of claim 1 does not constitute a technical scheme and does not belong to the object specified in Article 2, Paragraph 2 of the Patent Law.
In the reexamination procedure, the reexamination requestor added the feature "wherein, the occupied space of the neural network exceeds the occupied space threshold" to claim 1, and stated the following argument: the modified claim 1 improves the computer performance. Compared with running an uncompressed neural network, it will improve the running speed. And a person skilled in the art can appreciate that the performance of the electronic device (e.g. computer) must be improved due to the large remaining storage space and the fast running speed. It can be seen that the modified claim 1 improves the computer performance.
In view of the modifications and statements of the applicant for reconsideration, the examination division considers that: the compression process of the neural network in this application is a storage stage of the neural network model after the model has been constructed, and it is not a further improvement of the neural network algorithm itself, and the compression storage of the neural network is precisely used to improve the performance of the computer device itself, to improve the utilization rate of its storage space, and to improve the performance of the computer operation, which belongs to the performance improvement of the computer system itself; although the compression process of this application is a pruning operation on the neural network, its purpose is not to reduce the complexity of the algorithm, but to reduce the storage space of the neural network, and at the same time, reducing the complexity of the algorithm is not necessarily a rule of mental activities, and it needs to be determined according to whether its improvement method is related to the characteristics of the computer device and the data itself.
The panel decided to set aside the decision of the State Intellectual Property Office to reject this application.
Case 6: Method and system for optimizing computing resources of a convolutional neural network
Decision No. 233057 Decision Date 2020-11-05
Application number 201610779212.3; Application date 2016-08-30
The rejected decision in case 6 concerns claim 1 as follows:
A method for optimizing computing resources of a convolutional neural network, the特征在于, the method for optimizing computing resources of the convolutional neural network comprises the following steps:
The input Map of the convolutional neural network is split into a sub-Map matrix including a plurality of sub-Maps, where the size of the input Map is H×W, H>0, W>0, where the size of the input Map H×W is determined in advance according to factors such as image resolution, the size of the physical area that the image needs to reflect, and the size of the computing resources;
A convolution operation is performed separately on each of the所述 subMaps to obtain a computational result of a convolution operation on each of the所述 subMaps;
The calculation results of the convolution operation of all the所述子Map are in place spliced into the calculation results of the convolution operation of the所述input Map, where, in place splicing refers to placing the results of the convolution operation of the所述子Map at the corresponding position of the所述子Map in the said input Map.
It is pointed out in the decision to reject: claim 1 requests protection of a method for optimizing computing resources of a convolutional neural network, which solves the problems of large computing amount, long computing time, waste of computing resources and poor computing real-time of existing neural network, and the means it adopts is: to split the input Map of the convolutional neural network into a sub-Map matrix including a plurality of sub-Maps, and then perform separate convolution operations, and its means is an improvement in pure algorithm, and it does not adopt a technical means in accordance with natural laws. However, the effect of improving the computing efficiency of the algorithm, thus obtaining the effect of saving computing resources and improving real-time, is brought about by the improvement of the algorithm, and it is not a technical effect. Although the "size of the input Map H×W is determined in advance according to the image resolution, the size of the physical area that the image needs to reflect, and the size of the computing resources, etc." is mentioned in the scheme, the claim does not specifically reflect the combination of the neural network with the image field, let alone reflect the neural network solving what specific technical problems in the image. Therefore, claim 1 does not belong to a technical scheme and does not conform to the provisions of Article 2, Paragraph 2 of the Patent Law.
In the reexamination procedure, the reexamination applicant stated the following opinions: the problem to be solved by this application is to avoid the waste of computer resources and the poor real-time performance of the calculation, which belongs to the improvement of computer performance. Secondly, based on the pixel processing in the computer, the MAP is split using the pixel features in the image, so as to achieve the convolution operation of the MAP, which reflects the combination of pixels and neural networks, and solves the problem of waste of computer resources. Moreover, the convolutional neural network of this application can usually be applied in image processing technology, image recognition, natural language processing and physics and other technologies, and this application must be combined with the internal hardware of the computer to achieve the technical effects stated in this application.
In response to the statement of opinions made by the applicant for the review, the examination group believes that: each step in claim 1, as a whole, embodies how to use the convolutional neural network to process image data, which belongs to technical means. The problem to be solved is the problem of waste of computing resources and poor computing real-time when the layer间 re-use of the convolutional neural network is used to process image data. This problem is not only reflected in the problems of the convolutional neural network itself, but also in the problems of low computing efficiency and poor real-time performance when the convolutional neural network is applied to image processing, and how to improve the computing efficiency and real-time performance of image data processing reflects the problems that arise in the process of using natural laws, which belong to technical problems. The method of claim 1 achieves the effect of improving the computing efficiency and computing real-time of image processing by splitting the input Map of the image data and in-place splicing the computing results of each sub-Map, which belongs to technical effects. The size of the input Map of the convolutional neural network in the claims of this application is determined in advance according to the image resolution, the size of the physical area that the image needs to reflect, and the size of the computing resources and other factors. That is to say, it is clarified that the Maps processed in each step of the method are image data and how each step processes the image data, which reflects the close relationship between the convolutional neural network and the image data processing.
The panel decided to set aside the decision of the State Intellectual Property Office to reject this application.
Based on the re-examination results of Case 5 and Case 6, the following comparative analysis is made:
In the above two cases, the appellants in both cases stated in the opinions of the review procedure that the scheme to be protected could improve the internal performance of the computer system, and the review decisions of the review group were to overturn the rejection decision based on different reasons. Specifically:
In case 5, the applicant for reconsideration did not limit the specific technical field or fields in its authorized claim 1, which as a whole belongs to a scheme for compressing a general neural network model to optimize its structure; since the applicant for reconsideration limited the feature of "the occupied space of the neural network exceeds the occupied space threshold" in claim 1, this makes the structural optimization of the neural network closely linked to the storage performance of the computer device, and its model optimization effect also objectively achieves the performance improvement inside the computer system, so although the scheme does not limit any specific technical field, it does not affect its status as a technical scheme involving a computer program and constituting a patentable subject matter under the meaning of the Patent Law, and thus the examination division concluded that the decision to reject the application be revoked.
In case 6, although the applicant for reconsideration stated opinions similar to those in case 5 regarding the performance improvement inside the computer system, the examination board did not adopt the applicant's statement of opinions, but instead arrived at the conclusion of撤销the rejection decision based on the reasons for processing external technical data (image data) and the close relationship between the neural network and the processing of image data.
If the relevant scheme of the neural network can be closely related to the internal performance of the computer system (such as storage performance, running performance), then the patent application of the neural network can be obtained a wider protection range without limiting the specific technical field, and can be obtained as a general model algorithm. However, if the relevant scheme of the neural network can not be closely related to the internal performance of the computer system, the protection range of the patent application can be narrowed by limiting the specific technical field and the external technical data with precise technical meaning, so as to overcome the problem that it does not conform to the object of patent protection.
5. Conclusion
Based on the above analysis and comparison of the reexamination cases related to the object problem of neural networks, the following conclusion can be drawn:
1. Although neural networks can be used in a variety of different technical fields, a neural network itself does not constitute a technical scheme within the meaning of patent law and is not a protected object of a patent. If, in the statement of opinion or in the claims, a neural network is widely limited to a variety of different technical fields, it will, on the contrary, lead the examiner/panel to the conclusion that the overall scheme belongs to a general mathematical algorithm. Therefore, without other "patentable" basis to support, limiting the neural network to a variety of different technical fields as a general technical scheme is not enough to constitute a protected object of a patent in accordance with the provisions of patent law.
2、When a technical feature related to external technical data is recorded in the claim, which belongs to a specified technical field, it can only be determined that the scheme as a whole does not belong to the rules and methods of intellectual activities specified in Article 25 of the Patent Law based on this technical feature, but it is also necessary to further determine whether the scheme as a whole belongs to the technical scheme specified in Article 2, Paragraph 2 of the Patent Law. Specifically, it is necessary to analyze whether the relevant features of the neural network (such as the network layer or method step of the neural network) are closely related to the external technical data, so as to determine that the actual processing object of the neural network is the external technical data with specific technical meaning or the abstract general data. On the premise that there is no other "patentability" basis to support, if a neural network-related scheme does not reflect the close relationship between the relevant features of the neural network and the external technical data, the scheme is not sufficient to constitute a patent protection object in accordance with the provisions of the Patent Law.
3、If the relevant scheme of the neural network can be closely related to the internal performance of the computer system (such as storage performance, running performance), then the patent application of the neural network can be protected more widely without limiting the specific technical field, and can be obtained as a general model algorithm. However, if the relevant scheme of the neural network can not be closely related to the internal performance of the computer system, the protection scope of the patent application can be narrowed by limiting the specific technical field and the external technical data with precise technical meaning, so as to overcome the problem that it does not conform to the object of patent protection.