Computer-implemented method and apparatus for generating anomalies
By constructing two models to map and predict digital input, the system errors and abnormal problems existing in digital input automation processing in existing technologies are solved, and more stable and accurate processing effects are achieved.
Patent Information
- Application Number
- CN202410968545.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-07-18
- Filing Date
- 2024-07-18
- Publication Date
- 2025-01-21
AI Technical Summary
The prior art has system errors in the automated processing of digital inputs, resulting in undesirable exceptions.
By constructing two models: the first representation of the digital input in the state space in the first model is mapped to the digital input, the second model maps the digital input to the prediction, and determines the change of the first representative of the first indicator according to the measurement of the quality of the predicted quality. , Doward the prediction and cause abnormalities.
The production of abnormalities in the model is achieved, and the stability and accuracy of digital input automation processing are improved by identifying and using system errors.
Smart Images
Figure CN119337195A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a computer implemented method and apparatus for generating anomalies. Background Art
[0002] Models for automated processing of digital inputs are used in different fields. Such models may have systematic errors that lead to undesired anomalies. Summary of the invention
[0003] The device and the method according to the independent claims enable the generation of anomalies in a model.
[0004] The computer-implemented method provides for mapping a first representation of a digital input in a state space onto the digital input using a first model, mapping the digital input onto a prediction using a second model, the digital input being in particular a digital time series, a digital audio signal or a digital image, preferably a video image, a radar image, a laser radar (LiDAR) image, an ultrasound image, an image of a motion sensor or an infrared image, the first model being configured to map the first representation onto the digital input, the second model being configured to map the digital input onto a prediction for a content contained in the digital input, wherein a change in the first representation is determined according to a measure characterizing the quality of the prediction, the change deteriorating the prediction, wherein a second representation in the state space is determined according to the first representation and the change, and wherein an anomaly is generated according to the second representation. The first representation and the second representation are, for example, vectors in the state space. In the case of a digital image, the content is, for example, an object depicted in the digital image, such as a person or an object, such as a vehicle or a traffic sign. The first representation and the second representation define a concept that is independent of the content. In the case of a digital image, the concept is, for example, a weather condition or an exposure condition in which the digital image is detected. The second representation is determined by the method, the second representation generating the problematic input. The anomaly in the state space is thereby determined.
[0005] The first representation is, for example, randomly selected, or the first representation is determined based on the digital input using a third model that is constructed to map the digital input onto the first representation. By randomly selecting different first representations, different starting points for determining the second representation can be pre-specified. The corresponding starting point can be predicted by the third model. The third model is, for example, an inverse model of the first model.
[0006] The prediction preferably includes a classification of the content or a position of the content in the digital input. This means that the object is classified or the position of the object is determined.
[0007] In one example, the first representation represents a semantic concept about the content contained in the digital input. For example, the semantic concept describes weather conditions or exposure in a digital image. This means that using the first model, different semantic concepts can be mapped to corresponding digital inputs.
[0008] In one example, the first model is constructed to map a description and a first representation of content to be provided in the digital input to the digital input, wherein the description and the first representation of content to be provided in the digital input are mapped to the digital input using the first model. This means that different descriptions can be mapped to corresponding digital inputs using the first model.
[0009] It can be provided that the second representation is mapped onto a digital input, in particular a digital time series, a digital audio signal or a digital image, preferably a video image, a radar image, a lidar image, an ultrasound image, an image of a motion sensor or an infrared image, using the first model. This means that training data points are generated using the second representation, which training data points are expected to be problematic with respect to a correct prediction using the second model.
[0010] It can be provided that the first model is constructed to map a description of the content to be provided in the digital input and a second representation onto the digital input, wherein each digital input is determined for different descriptions according to the second representation using the first model, wherein a prediction is determined for each digital input using the second model, wherein it is checked whether at least part of the predictions have a common attribute, and wherein if the part of the predictions have a common attribute, a second representation suitable for generating an exception is identified. The second representation generates this attribute continuously. This means that the problem lies with the second model and not with the individual digital inputs. For example, in the case of digital images, the attribute is the similarity of the color of an object with a background color. For example, the attribute is the inability of the second model to make a prediction or a prediction with sufficient confidence for the digital input. For example, the inability is based on the fact that the content (e.g. the object) is unknown to the second model.
[0011] In one example, second representations are determined for different second models, wherein if the distance between the second representations determined for the different second models is less than a predetermined threshold, a second representation suitable for generating an anomaly can be identified. The distance can be a cosine distance between vectors depicting the second representation. Second models, especially second models of the same type, may have comparable problems. If the distance is small, the concept associated with the second representation is problematic for the second model. And if the distance is large, the concept is less problematic.
[0012] It may be provided that for identical descriptions of the content to be depicted and different second representations a frequency is determined, with which anomalies are detected, wherein the robustness of the second model with respect to the predicted changes is determined as a function of the frequency.
[0013] The method is preferably performed without human control.
[0014] The method preferably provides that a second model for controlling the technical system is formed as a function of the prediction, wherein the technical system is controlled as a function of the prediction of the second model.
[0015] The device for generating an exception comprises at least one processor and at least one memory, wherein the at least one processor is configured to execute instructions stored on the at least one memory, and when the instructions are executed by the at least one processor, the method is executed. The device has corresponding advantages to the method.
[0016] The computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method is executed. The computer program has advantages corresponding to the method. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Further advantageous embodiments can be found in the following description and in the accompanying drawings. In the drawings:
[0018] Figure 1 A schematic diagram showing an apparatus for generating an exception,
[0019] Figure 2 A schematic diagram showing a system for generating and processing images with the aid of machine learning,
[0020] Figure 3 A flow chart with steps in a method for generating an exception is shown. DETAILED DESCRIPTION
[0021] exist Figure 1 A device 100 for generating anomalies is schematically shown in FIG.
[0022] The device 100 includes at least one processor 102 and at least one memory 104. The at least one processor 102 is configured to execute instructions stored on the at least one memory 104, and when the instructions are executed by the at least one processor 102, the method described below for generating an exception runs.
[0023] A computer program may be provided, comprising instructions executable by at least one processor 102 .
[0024] The device 100 is configured to process a digital input 106. The digital input 106 includes content 108. The content 108 is, for example, a pedestrian, a vehicle or a road.
[0025] The digital input 106 is, for example, a digital image, in particular a video image, a radar image, a lidar image, an ultrasound image, an infrared image, an image of a motion sensor, etc. The digital input 106 is, for example, a digital time series or a digital audio signal.
[0026] It may be provided that the digital input 106 is detected by means of a sensor 110. The sensor 110 is, for example, a camera, a radar sensor, a lidar sensor, an ultrasonic sensor, an infrared sensor or a microphone. It may be provided that the sensor 110 is designed to detect a digital time series of a physical variable, in particular a speed, an acceleration, a yaw rate or a rotation angle. It may be provided that the device 100 includes the sensor 110.
[0027] In this example, device 100 is designed to operate a technical system 112 .
[0028] It may be provided that the device 100 is designed to control the actuator 114 using a signal that is dependent on the digital input 106. It may be provided that the device comprises the actuator 114.
[0029] Technical systems 112 are physical systems, such as robots, in particular vehicles, tools, domestic appliances, medical devices, personal assistance systems or access control systems.
[0030] Figure 2 A schematic diagram of a system 200 for generating and processing images with the aid of machine learning is shown.
[0031] The system 200 includes a first model 202 for generating an image and a second model 204 for processing an image. For example, the first model 202 is a deep neural network. The second model 204 is, for example, a deep neural network.
[0032] The first model 202 is configured to map a first representation 206 of the digital input 106 from a state space 208 and a description 210 of the content 108 to be provided in the digital input 106 onto the digital input 106 .
[0033] The first model 202 is, for example, a generator that is configured to map the semantic concepts and descriptions 210 encoded by the first representation 206 onto the digital input 106 .
[0034] For example, the digital input 106 is a digital image. The content 108 is, for example, a pedestrian, a vehicle, or a road. The semantic concept is, for example, weather conditions or light exposure.
[0035] The second model 204 is configured to map the digital input 106 onto the prediction 212. The system 200 is configured to determine a change 214 of the first representation 206 based on the prediction 212 and a measure m of quality of the prediction 212, the change degrading the prediction 212.
[0036] Second model 204 is, for example, a classifier, wherein prediction 212 includes a category of content 108 contained in digital input 106 .
[0037] Second model 204 is designed, for example, for semantic segmentation, wherein prediction 212 includes a semantic segmentation of content 108 contained in digital input 106 .
[0038] Second model 204 is designed, for example, for object recognition, wherein prediction 212 includes the position of content 108 (ie, the object) contained in digital input 106 .
[0039] In this example, the instructions define a system 200 for machine learning.
[0040] exist Figure 3 A flow chart with steps in a method for generating an exception is shown in The method is performed automatically, ie without manual control.
[0041] The method is based on the system for machine learning 200. The method is described by taking the generation of anomalies in the second model 204 as an example.
[0042] The method includes step 302 .
[0043] In step 302 , a first representation 206 is predefined.
[0044] In this example, the first representation 206 is randomly selected.
[0045] The first representation 206 represents, for example, a semantic concept regarding the content 108 contained in the digital input 106 .
[0046] In this example, a description 210 of content 108 to be provided in digital input 106 is predetermined.
[0047] Depending on the optimization method used, it is possible that the predetermined first representation 206 leads to a local optimum. Therefore, provision can be made for the method to be carried out for various predetermined first representations 206, wherein the second representation leading to the greatest degradation of the prediction 212 is determined.
[0048] For example, this situation can be avoided by using a third model that approximates the inverse function of the first model 202. For example, the third model is a deep neural network. For example, a style encoder network that approximates the inverse function of the generator is provided. It can be provided that the digital input 106 is given in advance according to the training data, and the first representation 206 is determined using the third model. For example, the third model is constructed to map the digital input 106 to the first representation 206. For example, the first representation 206 is determined according to the digital input 106 using the third model.
[0049] Step 304 is then executed.
[0050] In step 304 , first representation 206 of digital input 106 is mapped onto digital input 106 using first model 202 .
[0051] In this example, a description 210 and a first representation 206 of the content 108 to be provided in the digital input 106 are mapped onto the digital input 106 .
[0052] Step 306 is then executed.
[0053] In step 306 , the digital input 106 is mapped onto the prediction 212 using the second model 204 .
[0054] Prediction 212 may include a classification of content 108 or a location of content 108 within digital input 106 .
[0055] Step 308 is then executed.
[0056] In step 308 , a second representation in state space 208 is determined based on a metric characterizing the quality of prediction 212 , the second representation degrading prediction 212 .
[0057] For example, the second representation v is determined as:
[0058]
[0059] in
[0060] G denotes the first model 202, M denotes the second model 204, m denotes a metric, s denotes a description 210 from a set of descriptions S, and S and E denote expected values.
[0061] In this example, the metric m is selected to match the prediction 212 used, ie for example the metric m for classification is the classification accuracy, for semantic segmentation the metric is the Intersection over Union (IoU), and for object recognition the metric is the regression loss.
[0062] Preferably, the first model 202 and the second model 204 are differentiable. The second representation v is determined, for example, by a gradient descent method.
[0063] Step 310 is then performed.
[0064] In step 310, an exception is generated according to the second representation v.
[0065] For example, a respective digital input is determined for different descriptions according to the second representation v using the first model 202. For example, for each digital input, a prediction 212 is determined using the second model 204. In this example, it is checked whether at least part of the predictions have a common attribute, and if the part of the predictions have a common attribute, a second representation v suitable for generating an exception is identified.
[0066] For example, second representations v are determined for different second models, and if the distance between the second representations determined for the different second models is less than a predetermined threshold, a second representation v suitable for generating an anomaly is identified.
[0067] For example, for the same description of the content to be depicted and a different second representation v, the frequency of generating anomalies is determined. For example, the robustness of the second model 204 with respect to changes in the prediction 212 is determined based on the frequency.
[0068] For example, if the frequency is less than a predetermined threshold, robustness is identified.
[0069] It may be provided that the second representation v is mapped together with the first model 202 onto a digital input assigned to the second representation v.
[0070] In this example, the digital input is of the same type as the digital input 106 assigned to the first representation 206, in particular a digital time series, a digital audio signal or a digital image, preferably a video image, a radar image, a lidar image, an ultrasound image, an image of a motion sensor or an infrared image.
[0071] This digital input may supplement the training data with training data points that are known to be problematic for second model 204 .
[0072] Provision may be made to determine a plurality of second representations v. Provision may be made to determine a plurality of digital inputs for these second representations v. The training data is thereby supplemented with a large number of training data points which are problematic for the second model 204 .
[0073] Provision may be made to train the second model 204 using these additional training data points.
[0074] In one example, the digital input 106 represents a state of the technical system 112 or an environment of the technical system 112. For example, a digital image, a digital time series or a digital audio signal represents the state. This means that the second model 204 determines the prediction 212 based on the state.
[0075] Second model 204 is designed, for example, to actuate technical system 112 as a function of prediction 212 for digital input 106 . This means, for example, that second model 204 is designed to determine a signal for actuating an actuator, ie, for actuating technical system 112 , as a function of prediction 212 .
[0076] For example, a training is provided in which additional anomalies are generated as training data for second model 204. It can be provided that second model 204 is trained with anomalies and thus made more robust. As a result, anomalies that did not occur previously in the training are also less problematic for prediction 212 and for the control of technical system 112.
Claims
1. A computer-implemented method for generating an exception, characterized in that A first representation (206) of a digital input (106) in a state space (208) is mapped onto the digital input (106) using a first model (202) (304), and the digital input (106) is mapped onto a prediction (212) (306) using a second model (204), the digital input (106) being in particular a digital time series, a digital audio signal or a digital image, preferably a video image, a radar image, a lidar image, an ultrasound image, an image of a motion sensor or an infrared image, the first model being constructed to map the first representation onto the digital input (106), the second model being constructed to map the digital input (106) onto a prediction (212) for a content contained in the digital input (106), wherein the second representation (308) in the state space (208) is determined based on a metric characterizing the quality of the prediction (212), the second representation degrades the prediction (212), and wherein an anomaly (310) is generated based on the second representation.
2. The method according to claim 1, characterized in that The first representation (206) (302) is selected randomly or determined based on the digital input using a third model configured to map the digital input (106) to the first representation (206) (302).
3. The method according to any one of the preceding claims, characterized in that The prediction (212) includes a classification of the content (108) or includes a location of the content (108) in the digital input (106).
4. The method according to any one of the preceding claims, characterized in that The first representation (206) represents a semantic concept about content (108) contained in the digital input (106).
5. The method according to any one of the preceding claims, characterized in that The first model (202) is constructed to map a description (210) of content (108) to be provided in the digital input (106) and a first representation (206) to the digital input (106), wherein the description (210) of content (108) to be provided in the digital input (106) and the first representation (206) are mapped to the digital input (106) using the first model (202).
6. The method according to any one of the preceding claims, characterized in that The second representation is mapped onto a data input, in particular a digital time series, a digital audio signal or a digital image, preferably a video image, a radar image, a lidar image, an ultrasound image, an image of a motion sensor or an infrared image (310) using the first model (202).
7. The method according to any one of the preceding claims, characterized in that The first model (202) is configured to map a description of content (108) to be provided in the digital input (106) and the second representation onto the digital input, wherein one digital input is determined for each different description based on the second representation using the first model (202), wherein a prediction (212) is determined for each digital input using the second model (204), wherein it is checked whether at least some of the predictions have common properties, and wherein if the some of the predictions have common properties, a second representation (310) suitable for generating an exception is identified.
8. The method according to any one of the preceding claims, characterized in that Second representations are determined for different second models, wherein if a distance between the second representations determined for the different second models is less than a predetermined threshold, then a second representation suitable for generating an anomaly is identified (310).
9. The method according to any one of claims 7 or 8, characterized in that A frequency of generating anomalies (310) is determined for the same description of the content to be depicted and a different second representation, wherein a robustness of the second model (204) with respect to changes in the prediction (212) is determined based on the frequency.
10. The method according to any one of the preceding claims, characterized in that The method is performed without human control.
11. The method according to any one of the preceding claims, characterized in that The second model (204) is designed to control the technical system (112) as a function of the prediction (212), wherein the technical system (112) is controlled as a function of the prediction (212) of the second model (204).
12. A device (100) for generating an anomaly, characterized in that The device (100) comprises at least one processor (102) and at least one memory (104), wherein the at least one processor (102) is configured to execute instructions stored in the at least one memory (104), and when the instructions are executed by the at least one processor (102), the method according to any one of claims 1 to 11 is executed.
13. A computer program, characterized in that The computer program comprises instructions executable by at least one processor (102), and when the instructions are executed by the at least one processor (102), the method according to any one of claims 1 to 11 is executed.