HOME > English > Insights > Patent

Feng LU
Chinese Patent Attorney

 
I. Introduction
 
As the digital era commences, artificial intelligence is in widely applied to various fields including home furnishing, manufacture, finance, medical care, security, transportation, retail, education, and logistics, resulting in an increasing number of AI-related patent applications. The statistics show that the number of patent applications in the field of artificial intelligence in 2021 reached 30 times the number in 2015, more than half of which are filed in China. AI-related patents involve algorithm features; but the algorithm features per se are not technical features. This leads to a difficulty of how to consider algorithm features in determining the inventiveness of AI-related patents. Subsequently this article explores the criteria of inventiveness determination for AI-related patents with reference to a recent case of invalidation.
 
II. Provisions of the Guidelines for Patent Examination
 
The Guidelines for Patent Examination provides in Section 6.1.3 of Chapter 9, Part II:
 
In examining inventiveness of an application for an invention patent which contains both technical features and features of algorithms or business rules and methods, the examiner should consider the technical features and the features of algorithms or business rules and methods functionally in mutual support and interaction with the technical features as a whole. The phrase “functionally in mutual support and interaction” means that the features of algorithms or business rules and methods are closely joined with the technical features to form a technical means for solving a certain technical problem and achieves a corresponding technical effect.
 
For example, when applied to a specific technical field, if the algorithm according to the claim can solve a specific technical problem, the algorithm feature can be considered being functionally in mutual support and interaction with the technical feature and constituting a part of the technical means. In such a case, the contribution made by the algorithm feature to the technical solution should be taken into consideration in the inventiveness determination.
 
Accordingly, if an algorithm feature is “functionally in mutual support and interaction” with a technical feature, the algorithm feature and the technical feature should be considered as a whole in the inventiveness determination. However, it is not easy to determine that the algorithm feature and the technical feature are “functionally in mutual support and interaction”. The Guidelines for Patent Examination currently in effect provides a specific situation where the algorithm of the claim can solve a specific technical problem when applied to a specific technical field, the algorithm feature can be considered being functionally in mutual support and interaction with the technical feature. However, the provision “the algorithm of the claim can solve a specific technical problem when applied to a specific technical field” is not specific enough.
 
As shown above, there is a need to clarify the criteria of inventiveness determination for AI-related patents.
 
III. Case Summary
 
The subject patent is the patent No. ZL201910958076.6 of which the title is “Method for establishing scrap steel grading neural network model”.
The subject patent relates to application of AI technology in the steel industry. By feature extraction and deep learning for scrap steel grading classification based on the convolutional neural network technology, it realizes objective and accurate automatic classification scrap steel grading classification.
 
Claim 1 of the subject patent recites:
 
A method for establishing a scrap steel grading neural network model, the model configured for grading classification detection of scrap steel collection and storage, the method comprising: acquiring a plurality of images, determining different scrap steel grades of the plurality of images by visual inspection, preprocessing the images to remove invalid watermarks to improve image contrast, extracting image data characteristics of image data, and performing convolutional neural network learning on the extracted image data characteristics of different grades to form a grading neural network model with a grade classification output, wherein extraction of the image data characteristics is extraction implemented on a set obtained by carrying out convolutional neural network convolution calculation on pixelmatrix data of an image picture, and comprises: extracting object colors, edge features and texture features in an image and extracting associated features between object edges and textures in the image, wherein the object colors, the edge features and the texture features are formed by calculating a plurality of circuit convolution layers or convolution layers and pooling layers output by a set,
 
wherein a) extraction of object colors and edge features in the image is formed by a collection output calculated and output by three lines of convolutional layer plus pooling layer including, from a left side to a right side, a first line of one layer of pooling layer, a second line of two convolutional layers, and a third line of four convolutional layers; b) extraction of texture features in an image is extraction of the collection output of the extraction of object colors and edge features in the image, and is formed by a collection output calculated and output by three lines of convolutional layers including, from a left side to a right side, a first line of zero convolutional layer, a second line of two convolutional layers, and a third line of three convolutional layers; the texture features form an activation function (Relu activation) of the convolutional network;
 
a collection output calculated and output by at least three lines of convolutional layers or convolutional layers plus pooling layers forms extraction of object colors, edge features, and texture features in an image, each line having a different number of convolutional layers;
 
a number of lines of convolutional layers calculating the extraction of association features between edge and texture is larger than a number of lines of convolutional layers calculating the extraction of object colors, edge features, and texture features in an image.
 
Evidence 1 discloses the result of a test recognition performed using images of train wheels as material grains. The recognition result is the specific type of scrap steel.
 
Claim 1 of the subject patent is distinguished from Evidence 1 in the following aspects:
 
First, the application scenarios are different. Claim 1 is directed to a method for establishing a scrap steel grading neural network model, the model configured for grading classification detection of scrap steel collection and storage. The applications scenario of claim 1 is scrap steel grading. In contrast, Evidence 1 discloses a method for establishing a scrap steel type recognition neural network model and is applied to recognition of scrap steel types.
 
Second, the method steps are different. Claim 1 defines determining different scrap steel grades of a plurality of images by visual inspection in the step of acquiring images, extracting image data characteristics of different grades in the step of extracting image data characteristics, performing learning on the extracted image data characteristics of different grades and forming a grading neural network model with a grade classification output in the step of learning and training of the neural network model. Evidence 1 discloses the steps of acquiring images, pre-processing, feature extraction, and establishing a neural network model by training and learning. But the convolutional neural network model trained according to Evidence 1 is configured to recognize a specific type of scrap steel in an image of scrap steel, which is irrelevant to the grade of the scrap steel.
 
Third, the key parameters and the specific modular configurations are different. Claim 1 defines more specific contents for the extraction of image data characteristics, such as the parameters selected in feature extraction and the specific modular configurations for feature extraction. Evidence 1 does not contain such disclosure.
 
Based on the above distinguishing technical features, the technical problem to be actually solved by claim 1 is to establish a neural network model for grading scrap steel, so as to solve the problem of grading in the application scenario of grading and classification detection in scrap steel collection and storage, and to specifically determine the relevant data parameters and modules for solving the above problem.
 
The panel alleges that the entire Evidence 1 describes how to perform automatic recognition of the type of scrap steel; the method steps and the embodiments disclosed only involve how to perform type recognition and what type of material the recognition result is. Evidence 1 does not further describe or disclose how to grade the scrap steel. Therefore, there is no technical teaching for establishing a neural network model for grading scrap steel or grading scrap steel of various types which are mixed together in the application scenario, the steps, and the key parameters of Evidence 1.
 
Evidence 2 discloses the specific modular configurations for the extraction of image data characteristics by the convolutional neural network model as described in the third distinguishing aspect, and discloses that by the overall model architecture of Evidence 2, it is possible to speed up the training of the network and achieve more stable training.
 
However, the panel determines that Evidence 2 does not disclose which characteristics of the image data is extracted specifically, and does not disclose to which specific application scenario the extracted data characteristics are applied or the specific technical problem the extracted data characteristics solve in the application scenario; therefore, Evidence 2 does not provide the technical teaching for establishing a neural network model for grading scrap steel, not to mention teaching the specific parameters to be extracted for solving the existing technical problem.
 
Further, the panel determines that Evidence 3 does not disclose to which specific application scenario the extracted data characteristics are applied or the specific technical problem the extracted data characteristics solve in the application scenario, either; and thus, Evidence 3 does not provide the technical teaching for establishing a neural network model for grading scrap steel, not to mention teaching the specific parameters to be extracted for solving the existing technical problem.
 
In view of the foregoing, the panel determines that Claim 1 of the subject patent possesses inventiveness, and made the decision to maintain all claims of the subject patent to be valid.
 
IV. Case Analysis
 
In this case, both Claim 1 and Evidence 1 contain algorithm features. A difficulty of the case is how to consider these algorithm features in the determination of inventiveness. Although both Claim 1 and Evidence 1 relates to scrap steel, the application scenario of Claim 1 is grading the scrap steel while the application scenario of Evidence 1 is recognition of the type of the scrap steel. The difference in the application scenario leads to differences in the algorithm features, and substantive differences in particular. Therefore, the panel recognizes the positive effect of the algorithm features of claim 1 in the inventiveness determination.
 
This case is a typical case where algorithm features and technical features are functionally in mutual support and interaction, thereby providing significant reference for the inventiveness determination of AI-related patents. The CNIPA included this case in the Top Ten Cases of Patent Reexamination and Invalidation in 2022, stating that this case demonstrates refined criteria of inventiveness determination for AI-related patents including algorithm features and provides an example of the inventiveness determination for patents in the field of artificial intelligence. When artificial intelligence technology is involved, in the inventiveness determination for an invention patent including algorithm features, the algorithm and the application scenario should be considered as a whole; in particular, it should be taken into consideration whether or not the algorithm is, when applied in different scenarios, adapted in terms of the training mode, the key parameters, or the relevant steps, and whether or not the adaption of the algorithm solves a specific technical problem and achieves an advantageous technical effect.
 
On the other hand, assuming that the algorithm of Evidence 1 can be used in the application scenarios of the subject patent without making substantive adaption to the training mode, the key parameters, or the relevant steps, without any adaption for example or requiring only non-substantive and slight adaptions, the subject patent may not have inventiveness even based on considering the application scenario and the algorithm features as a whole.
 
V. Conclusion
 
As seen from this case, application scenario is critical to the grant of a patent relating to artificial intelligence. A pure algorithm for artificial intelligence without restriction of a specific application scenario would probably be directly determined to be ineligible for a patent under the Patent Law without going through inventiveness determination. The technical correlation between the application scenario and the algorithm features also plays an important role in the inventiveness determination. It is an urging demand that the criteria for the examination of AI-related patents would be refined and further clarified along with the development of the examination practice, so as to provide a direction for the drafting of the application documents for AI-related patents.
 
The foregoing is a summary of the experience of the author in the inventiveness determination for AI-related patents. Any comment or suggestion from the reader will be very welcome and much appreciated.