Method for making function of machine learning algorithm interpretable

Abstract

The invention relates to a method for making a function of a machine learning algorithm interpretable, the machine learning algorithm being designed to assign input data to one of at least two groups, and the method (1) having the following steps: providing the machine learning algorithm with input data (2); -assigning, by means of the machine learning algorithm, respective corresponding input data to one of the at least two groups for all of the provided input data (3); -selecting data (4) from a first of the at least two groups; -determining (5) data from a second group of said at least two groups that is most similar to the selected data among all data contained in said second group; -comparing (6) the selected data with the determined data in order to render the machine learning algorithm interpretable; and-providing a corresponding comparison result (7).

Description

Technical Field

[0001] The present invention relates to a method for making the functions of machine learning algorithms interpretable, by means of which the functions of machine learning algorithms can be made interpretable or understandable in a simple manner and with relatively low resource consumption. Background Art

[0002] Machine learning algorithms are based on using statistical methods to train a data processing system so that the data processing system can perform a specific task without having been explicitly programmed for that specific task initially. Here, the aim of machine learning is to build algorithms that can learn from data and make predictions. These algorithms create mathematical models, using which, for example, data can be classified.

[0003] Here, these machine learning algorithms include, for example, classification methods. A classification method is a method for describing the assignment or grouping of observations into predefined categories.

[0004] Such classification methods are applied, for example, in methods for identifying anomalies on the surface of products manufactured by a manufacturing process. An example of such a method is an automatic inspection, which is designed to identify defects in the corresponding products by means of image processing methods.

[0005] Here, it is generally desired to make the functions of machine learning algorithms interpretable or understandable, for example, in order to increase trust in the corresponding machine learning algorithms and / or to optimize the corresponding machine learning algorithms accordingly.

[0006] Here, for example, it is known to generate artificial counterfactual examples, based on which the functions of machine learning algorithms are to be made interpretable. However, the generation of artificial counterfactual examples involves relatively high resource consumption, such as a large consumption of memory and / or processor capacity. In addition, the quality of the generated artificial counterfactual examples is usually limited.

[0007] A method for generating counterfactual examples of a neural network is known from the published document EP 3796228 A1. Summary of the Invention

[0008] Therefore, the task underlying the present invention is to describe an improved method for making machine learning algorithms interpretable.

[0009] This task is solved by a method for making machine learning algorithms interpretable according to the features of patent independent claim 1.

[0010] In addition, this task is also solved by a system for making machine learning algorithms interpretable according to the features of patent independent claim 6.

[0011] According to one embodiment of the present invention, this task is solved by a method for making the function of a machine learning algorithm interpretable, wherein the machine learning algorithm is designed to assign input data to one of at least two groups respectively, and wherein the method has:

[0012] Providing input data for the machine learning algorithm;

[0013] Through the machine learning algorithm, for all of the provided input data, respectively assign the corresponding input data to one of the at least two groups;

[0014] Selecting data from the first group of the at least two groups;

[0015] Determining the data in the second group of the at least two groups that is most similar to the selected data among all the data included in the second group;

[0016] Comparing the selected data with the determined data in order to make the machine learning algorithm interpretable; and

[0017] Providing a corresponding comparison result.

[0018] In this case, the input data is understood as the data that is assigned to the corresponding output data or output value by the machine learning algorithm, that is, especially the data of one of the at least two groups, especially the corresponding sensor data.

[0019] Therefore, a method for making the function of a machine learning algorithm interpretable is described, which is based on: determining the data that is most similar to each other from respectively different groups, wherein by comparing these data, the differences in the assignment of the data to the corresponding groups can be indicated.

[0020] Therefore, a method for making the function of a machine learning algorithm interpretable is described, by using which high-quality counterfactual examples can be generated, however, this method does not require generating artificial counterfactual examples, that is, a method that can be used to make the function of a machine learning algorithm interpretable in a simple manner and with relatively low resource consumption is described.

[0021] Therefore, generally speaking, an improved method for making a machine learning algorithm interpretable is described.

[0022] In one embodiment, the step of determining the data in the second group that is most similar to the selected data herein has: applying at least one encoder.

[0023] Herein, an encoder or an autoencoder is a machine learning algorithm that is designed to extract specific features from data and make complex data understandable.

[0024] Thus, similar data can be determined based on known machine learning algorithms in a simple manner without complex and resource-intensive adjustments.

[0025] In addition, the method may further include: retraining the machine learning algorithm based on these comparison results. In particular, the weaknesses of the corresponding machine learning algorithm can be discovered and eliminated in a simple manner and with relatively low resource consumption.

[0026] These input data may also include sensor data.

[0027] A sensor, also known as a detector, (measurement parameter or measurement) recorder, or (measurement) probe, is a technical component that can qualitatively or quantitatively detect specific physical or chemical properties and / or material properties of the surrounding environment of the technical component as a measurement parameter.

[0028] Therefore, situations outside the data processing system on which the method is executed can be considered and incorporated into making the function of the machine learning algorithm interpretable.

[0029] For example, the machine learning algorithm may be a machine learning algorithm for automatically optically inspecting products manufactured by a manufacturing process, where the input data are image data of the products manufactured by the manufacturing process detected by sensors.

[0030] A manufacturing process is generally understood as a standardized workflow in which products are manufactured by machining and processing raw materials or intermediate products by machine and / or manually using specified manufacturing processes, working equipment, and production materials. According to these comparison results, the just-mentioned manufactured products can be discarded and thus not further processed, or the manufactured products can be released for subsequent processing steps.

[0031] Especially in the method of automatic optical inspection, it is important here to understand the working principle of these methods and thus make the working principle interpretable.

[0032] Using another embodiment of the present invention, a system for making the function of a machine learning algorithm interpretable is also described, where the machine learning algorithm is designed to assign input data to one of at least two groups respectively, and where the system has:

[0033] A first providing unit, which is designed to provide input data for the machine learning algorithm;

[0034] An assignment unit, which is designed to assign the corresponding input data to one of the at least two groups respectively for all of the provided input data through the machine learning algorithm;

[0035] A selection unit, which is designed to: select data from a first group of the at least two groups;

[0036] A determination unit, which is designed to: determine data that is most similar to the selected data among all the data included in a second group of the at least two groups;

[0037] A comparison unit, which is designed to: compare the selected data with the determined data so as to make the machine learning algorithm interpretable; and

[0038] A second providing unit, which is designed to: provide a corresponding comparison result.

[0039] Therefore, an improved system for making a machine learning algorithm interpretable is described. In particular, a system for making the function of a machine learning algorithm interpretable is described. Using this system, high-quality counterfactual examples can be generated. However, this system does not need to generate artificial counterfactual examples, that is, a system that can be used to make the function of a machine learning algorithm interpretable in a simple manner and with relatively low resource consumption is described.

[0040] In one embodiment, the determination unit is hereby designed to: apply at least one encoder so as to determine data. Therefore, similar data can be determined based on a known machine learning algorithm in a simple manner without the need for complex and resource-intensive adjustments.

[0041] In addition, the system may further have a retraining unit, which is designed to: retrain the machine learning algorithm based on these comparison results. In particular, the weaknesses of the corresponding machine learning algorithm can be discovered and eliminated in a simple manner and with relatively low resource consumption.

[0042] These input data may also have sensor data. Therefore, situations outside the data processing system on which the method is executed can be considered and incorporated into making the function of the machine learning algorithm interpretable.

[0043] For example, the machine learning algorithm may be a machine learning algorithm for automatically optically inspecting products manufactured by a manufacturing process, where the input data is image data of the products manufactured by the manufacturing process detected by sensors. Especially in the automatic optical inspection method, it is important here to: understand how these methods work and thus make the working principle interpretable.

[0044] Furthermore, using another embodiment of the present invention, a computer program is also described, which has program code for performing the above-mentioned method for making the functions of a machine learning algorithm interpretable when the computer program is executed on a computer.

[0045] Furthermore, using another embodiment of the present invention, a computer-readable data carrier is also described, which has the program code of a computer program for performing the above-mentioned method for providing training data to make the functions of a machine learning algorithm interpretable when the computer program is executed on a computer.

[0046] Herein, the computer program and the computer-readable data carrier each have the following advantages: The computer program and the computer-readable data carrier are designed to perform an improved method for making a machine learning algorithm interpretable. The computer program and the computer-readable data carrier are particularly designed to: perform a method for making the functions of a machine learning algorithm interpretable, by which high-quality counterfactual examples can be generated, however, the method does not require generating artificial counterfactual examples, that is, a method is described that can be used to make the functions of a machine learning algorithm interpretable in a simple manner and with relatively low resource consumption.

[0047] In summary, it should be emphasized that: using the present invention, a method for making the functions of a machine learning algorithm interpretable is described, by which the functions of a machine learning algorithm can be made interpretable in a simple manner and with relatively low resource consumption.

[0048] The described design solutions and extensions can be combined with each other arbitrarily.

[0049] Other possible design solutions, extensions, and implementation solutions of the present invention also include combinations of features not explicitly mentioned before or below in the description of the embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are intended to provide a further understanding of the embodiments of the present invention. These drawings illustrate the embodiments and are used in combination with the description to explain the principles and designs of the present invention.

[0051] Multiple advantages of other embodiments and the advantages mentioned are derived with reference to these drawings. The elements presented in these drawings are not necessarily shown to scale with each other.

[0052] Wherein:

[0053] Figure 1 A flowchart of a method for making the functions of a machine learning algorithm interpretable according to an embodiment of the present invention is shown;

[0054] Figure 2 FIG. 1 shows a schematic block diagram of a system for making the function of a machine learning algorithm interpretable according to an embodiment of the present invention.

[0055] In the figures of these drawings, unless otherwise stated, the same reference numerals denote the same or functionally identical elements, components or assemblies. DETAILED DESCRIPTION

[0056] Figure 1 FIG. 2 shows a flowchart of Method 1 for determining at least one change - point in a time series of sensor values according to an embodiment of the present invention.

[0057] Here, these machine learning algorithms include, for example, classification methods. A classification method is a method that describes the assignment or grouping of observations into predefined classes.

[0058] Such a classification method is applied, for example, in a method for identifying anomalies on the surface of a product manufactured by a manufacturing process. An example of such a method is an automatic inspection, which is designed to identify defects in the corresponding product by means of an image - processing method.

[0059] Here, it is generally desired to make the function of a machine learning algorithm interpretable or understandable, for example, in order to increase trust in the corresponding machine learning algorithm and / or optimize the corresponding machine learning algorithm accordingly.

[0060] Here, for example, it is known that artificial counterfactual examples are generated, and the function of the machine learning algorithm is to be made interpretable based on these artificial counterfactual examples. However, the generation of artificial counterfactual examples involves a relatively high resource consumption, such as a large consumption of memory and / or processor capacity. In addition, the quality of the generated artificial counterfactual examples is usually limited.

[0061] Here Figure 1 FIG. 3 shows a Method 1 for making the function of a machine learning algorithm interpretable, wherein the machine learning algorithm is designed to assign input data to one of at least two groups respectively, and wherein Method 1 has:

[0062] Step 2: Provide input data for the machine learning algorithm;

[0063] Step 3: Through the machine learning algorithm, for all of the provided input data, assign the corresponding input data to one of the at least two groups respectively;

[0064] Step 4: Select data from the first of the at least two groups;

[0065] Step 5: Determine the data in the second group of the at least two groups that is most similar to the selected data among all the data included in the second group;

[0066] Step 6: Compare the selected data with the determined data so that the machine learning algorithm is interpretable; and

[0067] Step 7: Provide the corresponding comparison result.

[0068] Therefore, a method 1 for interpreting the function of a machine learning algorithm is described. Using this method, high-quality counterfactual examples can be generated. However, this method does not require generating artificial counterfactual examples, that is, a method is described that can be used to interpret the function of a machine learning algorithm in a simple manner and with relatively low resource consumption.

[0069] Therefore, generally, an improved method 1 for interpreting a machine learning algorithm is described.

[0070] Figure 1 In particular, method 1 is shown, in which counterfactual examples are generated by determining the data in one category that is most similar to the selected data in another category.

[0071] Here, step 4 of selecting data from the first group of the at least two groups may have: randomly selecting data; or selecting data based on corresponding specifications, such as application-specific specifications.

[0072] According to Figure 1 the embodiment, step of determining the data in the second group that is most similar to the selected data here has: applying at least one encoder.

[0073] In particular, here, feature vectors can be formed respectively based on the selected data and all elements of the second group, and data similar to the selected data can be determined based on these feature vectors.

[0074] As Figure 1 shown, method 1 also has step 8: Retrain the machine learning algorithm based on these comparison results.

[0075] Here, method 1 is particularly designed to: evaluate the model quality; and correspondingly optimize the machine learning algorithm.

[0076] In addition, these input data also have sensor data.

[0077] According to Figure 1An embodiment, in particular, a machine learning algorithm for automatically optically inspecting products manufactured by a manufacturing process, wherein the input data are image data of the products manufactured by the manufacturing process detected by a sensor.

[0078] Herein, the selected data may in particular be data that have been classified as abnormal (i.e., having errors) or having anomalies, and data classified as normal that are most similar to these data are determined.

[0079] Furthermore, based on the corresponding classification results, products classified as abnormal by the machine learning algorithm can also be automatically discarded herein.

[0080] Figure 2 A schematic block diagram of a system 10 for determining at least one change point in a time series of sensor values according to an embodiment of the present invention is shown.

[0081] Figure 2 In particular, a system 10 for making the function of a machine learning algorithm interpretable is shown, wherein the machine learning algorithm is designed to assign input data to one of at least two groups respectively, and wherein the system 10 has:

[0082] A first providing unit 11, which is designed to: provide input data for the machine learning algorithm;

[0083] An assignment unit 12, which is designed to: through the machine learning algorithm, assign the corresponding input data to one of the at least two groups respectively for all of the provided input data;

[0084] A selection unit 13, which is designed to: select data from a first group of the at least two groups;

[0085] A determination unit 14, which is designed to: determine data in a second group of the at least two groups that are most similar to the selected data among all the data included in the second group;

[0086] A comparison unit 15, which is designed to: compare the selected data with the determined data so as to make the machine learning algorithm interpretable; and

[0087] A second providing unit, which is designed to: provide the corresponding comparison result.

[0088] Here, the first providing unit can in particular be a receiver, which is designed to receive corresponding data, in particular sensor data. The second providing unit can further be a transmitter, which is designed to transmit corresponding information or data. In addition, the first providing unit and the second providing unit can also be integrated into a common transceiver here.

[0089] In addition, the allocation unit, the selection unit, the determination unit, and the comparison unit can for example be implemented respectively based on code stored in a memory and executable by a processor.

[0090] According to Figure 2 the embodiment, the determination unit 14 is also designed here to: apply at least one encoder in order to determine data.

[0091] As Figure 2 further shown, the system 10 also has a retraining unit 17, which is designed to: based on these comparison results, retrain the machine learning algorithm.

[0092] Here, the retraining unit can for example also be implemented based on code stored in a memory and executable by a processor.

[0093] In addition, these input data again have sensor data.

[0094] In particular, according to Figure 2 the embodiment, the machine learning algorithm is again a machine learning algorithm for automatically optically inspecting products manufactured by a manufacturing process, where these input data are image data of products manufactured by the manufacturing process detected by sensors.

[0095] In addition, the shown system 10 is designed to: execute the above-mentioned method for making the function of the machine learning algorithm interpretable.

Claims

1. A method for making a function of a machine learning algorithm interpretable, wherein: The machine learning algorithm is designed to assign input data to one of at least two groups, respectively, and wherein the method (1) comprises the following steps: - providing input data to the machine learning algorithm (2); - using the machine learning algorithm, for all the input data provided, respectively assigning the corresponding input data to one of the at least two groups (3); - selecting data from a first group of the at least two groups (4); - determining data from a second group of the at least two groups which is most similar to the selected data among all data contained in the second group (5); - comparing the selected data with the determined data in order to make the machine learning algorithm interpretable (6); and - Provide corresponding comparison results (7).

2. The method (1) according to claim 1, wherein: The step (5) of determining data from said second group which is most similar to the selected data comprises applying at least one encoder.

3. The method (1) according to claim 1 or 2, wherein: The method (1) further comprises the following steps: - Based on the comparison result, retraining the machine learning algorithm (8).

4. The method (1) according to any one of claims 1 to 3, wherein: The input data includes sensor data.

5. The method (1) according to claim 4, wherein: The machine learning algorithm is a machine learning algorithm for automatic optical inspection of products manufactured by a manufacturing process, and wherein the input data is image data of the products manufactured by the manufacturing process detected by a sensor.

6. A system for making a function of a machine learning algorithm interpretable, wherein: The machine learning algorithm is designed to assign input data to one of at least two groups respectively, and wherein the system (10) has: A first providing unit (11), the first providing unit being configured to: provide input data for the machine learning algorithm; An allocation unit (12), wherein the allocation unit is configured to: allocate corresponding input data to one of the at least two groups respectively for all the input data provided by the machine learning algorithm; A selection unit (13), the selection unit being designed to: select data from a first group of the at least two groups; A determination unit (14), the determination unit being configured to: determine data from a second group of the at least two groups that is most similar to the selected data among all data contained in the second group; a comparison unit (15), the comparison unit being designed to compare the selected data with the determined data in order to make the machine learning algorithm interpretable; and A second providing unit (16), wherein the second providing unit is configured to provide a corresponding comparison result.

7. The system (10) according to claim 6, wherein: The determination unit (14) is designed to use at least one encoder in order to determine the data.

8. The system (10) according to claim 6 or 7, wherein: The system (10) further comprises a retraining unit (17), which is designed to retrain the machine learning algorithm based on the comparison result.

9. A computer program having a program code for executing the method for making a function of a machine learning algorithm interpretable according to any one of claims 1 to 5 when the computer program is executed on a computer.

10. A computer-readable data carrier having a program code of a computer program in order to perform the method for providing training data to make the functions of a machine learning algorithm interpretable according to any one of claims 1 to 5 when the computer program is executed on a computer.

Images