Computer-implemented method and device for generating anomaly
The method addresses systematic errors in digital input models by using multiple models to characterize and mitigate abnormalities, enhancing prediction reliability and accuracy in digital input processing.
Patent Information
- Application Number
- JP2024114095
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-07-18
- Filing Date
- 2024-07-17
- Publication Date
- 2025-01-30
AI Technical Summary
Existing digital input models suffer from systematic errors that lead to unwanted abnormalities, particularly in processing digital time series, digital audio signals, digital images, and sensor data, which affect the reliability of predictions and classifications.
A method and device using multiple models to identify and mitigate abnormalities by mapping digital inputs to state spaces, employing a first expression and a second expression to characterize prediction quality, and a third model to optimize starting points, ensuring robustness and reliability of predictions.
The method enhances the reliability and robustness of predictions by identifying and addressing abnormalities in digital input models, improving the accuracy and consistency of predictions in technical systems.
Smart Images
Figure 2025015477000001_ABST
Abstract
Description
[Technical field]
[0001] Background technology Models for the automatic processing of digital inputs are used in many different fields. Such models can have systematic errors that result in undesirable anomalies. Summary of the Invention [Problem to be solved by the invention]
[0002] Disclosure of the Invention The device and the method according to the independent claims allow the generation of anomalies in the model. [Means for solving the problem]
[0003] It is envisaged in the computer-implemented method to map a first representation of a digital input in a state space, 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, to the digital input using a first model configured to map the first representation to the digital input, to map the digital input to a prediction of a content contained in the digital input using a second model configured to map the digital input to this prediction, to identify a change in the first representation that deteriorates this prediction depending on a measure characterizing the quality of the prediction, to identify a second representation in the state space depending on the first representation and this change, and to generate an anomaly depending on the second representation. The first and second representations are for example vectors in the state space. The content is for example an object, such as a human being, or an object, such as a vehicle or a traffic sign, displayed in the digital image in the case of a digital image. The first and second representations define a content-independent concept. In the case of a digital image, the concept is, for example, the weather conditions or exposure conditions under which the digital image was found. The method identifies second representations that generate problematic inputs. This identifies anomalies in the state space.
[0004] The first representation is determined depending on the digital input, for example chosen randomly or using a third model configured to map the digital input to the first representation. By randomly selecting different first representations, different starting points can be set for the determination of the second representation. Each starting point can be predicted by the third model. The third model is, for example, an inverse model of the first model.
[0005] The prediction preferably involves a classification of the content or a location of the content in the digital input, which means that an object is classified or a location of an object is identified.
[0006] In one example, the first representation represents a semantic concept for the content contained in the digital input, for example describing weather conditions or exposure conditions in a digital image. This means that different semantic concepts can be mapped to each digital input by the first model.
[0007] In one example, the first model is configured to map a description of the content to be provided in the digital input and a first representation to the digital input, the first model maps a description of the content to be provided in the digital input and a first representation to the digital input, which means that different descriptions can be mapped to each digital input by the first model.
[0008] It may be envisaged that the first model maps the second representation 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, which means that the second representation generates training data points that may be problematic for a correct prediction by the second model.
[0009] It may be assumed that the first model is configured to map a description of the content to be provided in the digital input and a second representation to the digital input, where the first model identifies one digital input each depending on the second representation for different descriptions, and for each digital input, the second model identifies a prediction, checks whether at least some of the predictions have a common characteristic, and if some of the predictions have a common characteristic, identifies a second representation suitable for generating an anomaly. The second representation consistently generates this characteristic. This means that the problem lies in the second model and not in the individual digital input. The characteristic is, for example, in the case of a digital image, the similarity of the color of the object and the color of the background. The characteristic is, for example, the inability of the second model to make a prediction for the digital input or to make a prediction with sufficient confidence. This inability is, for example, based on the content, for example the object, being unknown to the second model.
[0010] In one example, second representations are identified for different second models, and second representations suitable for generating anomalies are identified if the interval between the second representations identified for the different second models is smaller than a predefined threshold. This interval may be the cosine interval between the vectors representing the second representations. Second models, particularly second models of the same type, may have comparable problems in some cases. If this interval is short, the concepts associated with these second representations are problematic for the second models. Conversely, if this interval is long, the concepts are less problematic.
[0011] It may be assumed that for the same description and different second representations of the content to be displayed, the frequency with which anomalies are identified is determined, and in this case the robustness of the second model to changes in prediction is determined depending on this frequency.
[0012] The method is preferably performed without human control.
[0013] In the method, it is preferably provided that the second model is configured to control the technical system depending on the prediction, in which case the technical system is controlled depending on the prediction of the second model.
[0014] An apparatus for generating anomalies includes at least one processor and at least one memory, the at least one processor configured to execute instructions stored in the at least one memory that, when executed by the at least one processor, implements the method. The apparatus has advantages corresponding to those of the method.
[0015] The computer program includes instructions that can be executed by at least one processor, and when the instructions are executed by the at least one processor, the method is performed. The computer program has advantages corresponding to those of the method.
[0016] Further advantageous embodiments can be seen from the following description and the drawings. [Brief description of the drawings]
[0017] [Figure 1] FIG. 1 is a schematic diagram of an apparatus for generating anomalies. [Diagram 2] FIG. 1 is a schematic diagram of a system for generating and processing images using machine learning. [Diagram 3] 1 is a flow chart with steps in a method for generating anomalies. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0018] FIG. 1 shows a schematic of an apparatus 100 for generating anomalies.
[0019] The apparatus 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 in the at least one memory 104 that, when executed by the at least one processor 102, perform a method for generating anomalies, described below.
[0020] A computer program may be provided that includes instructions executable by at least one processor 102 .
[0021] The device 100 is configured to process a digital input 106. The digital input 106 includes content 108. The content 108 may be, for example, a pedestrian, a vehicle, or a road.
[0022] 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.
[0023] It may be assumed that the digital input 106 is detected by 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 assumed that the sensor 110 is configured to detect a digital time series of a physical quantity, in particular a speed, an acceleration, a yaw rate or a rotation angle. It may be assumed that the device 100 includes the sensor 110.
[0024] The device 100 is in this example configured to drive and control a technical system 112 .
[0025] It may be assumed that the device 100 is configured to control an actuator 114 by a signal dependent on the digital input 106. It may be assumed that the device includes the actuator 114.
[0026] The technical system 112 is a physical system, for example a robot, in particular a vehicle, a tool, a domestic device, a medical device, a personal assistance system or an access control system.
[0027] FIG. 2 shows a schematic diagram of a system 200 for generating and processing images using machine learning.
[0028] The system 200 includes a first model 202 for generating an image and a second model 204 for processing an image. The first model 202 is, for example, a deep neural network. The second model 204 is, for example, a deep neural network.
[0029] 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 .
[0030] The first model 202 is, for example, a generator configured to map the semantic concepts and descriptions 210 encoded by the first representation 206 to the digital input 106 .
[0031] 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 a weather condition or an exposure situation.
[0032] The second model 204 is configured to map the digital input 106 to a prediction 212. The system 200 is configured to identify changes 214 in the first representation 206 that deteriorate the prediction 212 depending on the prediction 212 and a measure m for the quality of the prediction 212.
[0033] The second model 204 is, for example, a classifier, where the predictions 212 include the class of the content 108 contained in the digital input 106 .
[0034] The second model 204 is configured, for example, for semantic segmentation, in which case the predictions 212 include semantic segmentations for the content 108 included in the digital input 106 .
[0035] The second model 204 is configured, for example, for object identification, where the prediction 212 includes the content 108 contained in the digital input 106, i.e. the location of the object.
[0036] The instructions, in this example, define a system 200 for machine learning.
[0037] A flow chart with steps in a method for generating anomalies is shown in Figure 3. The method is performed automatically, i.e. without human control.
[0038] The method is based on a system for machine learning 200. The method is illustrated with respect to the example of generating anomalies in a second model 204.
[0039] The method includes step 302.
[0040] In step 302, the first representation 206 is established.
[0041] The first representation 206 is selected randomly in this example.
[0042] The first representation 206 represents a semantic concept for the content 108 contained in the digital input 106, for example.
[0043] In this example, a description 210 of the content 108 to be provided in the digital input 106 is provided.
[0044] Depending on the optimization method used, the set first representation 206 may lead to a local optimum. It may therefore be envisaged to carry out the method for different set first representations 206, in which case the second representation that leads to the greatest deterioration of the prediction 212 is identified.
[0045] This can be avoided by a third model, for example approximating the inverse of the first model 202. The third model is for example a deep neural network. For example, a style encoder network is provided that approximates the inverse of the generator. It may be assumed that the digital input 106 is set from training data and the first representation 206 is determined using the third model. The third model is for example configured to map the digital input 106 to the first representation 206. The first representation 206 is determined depending on the digital input 106, for example, using the third model.
[0046] Then, step 304 is executed.
[0047] In step 304 , the first representation 206 of the digital input 106 is mapped to the digital input 106 by the first model 202 .
[0048] 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 to the digital input 106 .
[0049] Then, step 306 is executed.
[0050] In step 306 , the digital input 106 is mapped to a prediction 212 by the second model 204 .
[0051] The prediction 212 may include a classification of the content 108 or a location of the content 108 in the digital input 106 .
[0052] Then, step 308 is executed.
[0053] In step 308, second representations in the state space 208 that make the prediction 212 worse are identified depending on the measure that characterizes the quality of the prediction 212.
[0054] For example, a second expression v may be specified as follows:
number
[0055] The measure m is chosen in this example to suit the prediction 212 used, i.e., for example, measure m is the classification accuracy in the case of classification, the Intersection over Union (IoU) in the case of semantic segmentation, and the regression loss in the case of object identification.
[0056] Preferably, the first model 202 and the second model 204 are distinguishable. The second representation v is identified, for example, by gradient descent.
[0057] Then, step 310 is executed.
[0058] In step 310, anomalies are generated depending on the second representation v.
[0059] For example, the first model 202 identifies one digital input each for different descriptions depending on the second representation v. For each digital input, a prediction 212 is identified, for example, by the second model 204. In this example, it is checked whether at least some of the predictions have common properties, and if some of the predictions do, a suitable second representation v for generating an anomaly is identified.
[0060] For example, a second representation v is identified for different second models, and a second representation v suitable for generating an anomaly is identified when the spacing between the second representations identified for the different second models is less than a predetermined threshold.
[0061] For example, for the same description of the content to be displayed and different second representations v, the frequency with which anomalies are generated is determined. The robustness of the second model 204 to changes in the prediction 212 is determined, for example, depending on this frequency.
[0062] Robustness is identified, for example, if the frequency is below a predefined threshold.
[0063] It may be assumed that the second representation v is mapped by the first model 202 to a digital input that is assigned to the second representation v.
[0064] This digital input is in this example 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, a motion sensor image or an infrared image.
[0065] This digital input may supplement the training data with training data points that are already known to be problematic for the second model 204 .
[0066] It may be envisaged to identify a number of second representations v. It may be envisaged to identify a number of digital inputs for these second representations v. This supplements the training data with a number of training data points that are problematic for the second model 204.
[0067] The second model 204 may be envisioned to be trained with these additional training data points.
[0068] In one example, the digital input 106 characterizes a state of the technical system 112 or a state of the environment of the technical system 112. A digital image, a digital time series or a digital audio signal characterize for example a state. This means that the second model 204 specifies a prediction 212 depending on the state.
[0069] The second model 204 is configured for example to control the technical system 112 depending on a prediction 212 of the digital input 106. This means, for example, that the second model 204 is configured for determining signals for controlling actuators, i.e. for controlling the technical system 112, depending on the prediction 212.
[0070] For example, a training is provided in which additional anomalies are generated as training data for the second model 204. It may be envisaged to train the second model 204 with these anomalies and thus make it more robust, so that anomalies that did not occur beforehand during training are subsequently less problematic for the prediction 212 and the driving control of the technical system 112.
Claims
1. 1. A computer-implemented method for generating anomalies, comprising: Mapping (304) a first representation (206) of a digital input (106) in a state space (208), in particular a first representation (206) of 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, onto the digital input (106) using a first model (202) configured to map the first representation onto the digital input (106); Mapping (306) the digital input (106) to a prediction (212) of content contained in the digital input (106) using a second model (204) configured to map the digital input (106) to the prediction (212); identifying (308) a second representation in the state space (208) that deteriorates the prediction (212) in dependence on a measure characterizing the quality of the prediction (212); generating anomalies (310) dependent on the second representation; 13. A computer-implemented method comprising:
2. The first representation (206) is Randomly select (302), or 2. The method of claim 1, further comprising: determining (302) the digital input dependent on the third model configured to map the digital input to the first representation (206).
3. The method of claim 1 or 2, wherein the prediction (212) comprises a classification of the content (108) or a location of the content (108) in the digital input (106).
4. The method of claim 1 , wherein the first representation (206) represents a semantic concept for content (108) contained in the digital input (106).
5. 5. The method of claim 1, wherein the first model is configured to map a description of the content to be provided in the digital input and the first representation to the digital input, the description of the content to be provided in the digital input and the first representation being mapped to the digital input by the first model.
6. 6. The method according to claim 1, wherein the second representation is mapped (310) by the first model (202) to 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, a motion sensor image or an infrared image.
7. 7. The method of claim 1, wherein the first model (202) is configured to map descriptions of the content (108) to be provided in the digital input (106) and the second representation to the digital input, the first model (202) identifying a respective digital input depending on the second representation for different descriptions, and the second model (204) identifying the predictions (212) for each digital input, checking whether at least some of the predictions have common characteristics, and identifying (310) a second representation suitable for generating the anomaly if some of the predictions have common characteristics.
8. 8. The method according to claim 1, further comprising identifying (310) the second representation for different second models, and identifying a suitable second representation for generating the anomaly if a spacing between the second representations identified for the different second models is smaller than a predefined threshold.
9. 9. The method according to claim 7 or 8, further comprising determining a frequency with which anomalies are identified (310) for an identical description and a different second representation of the content to be displayed, and determining a robustness of the second model (204) to changes in the prediction (212) in dependence on the frequency.
10. 10. The method of any one of claims 1 to 9, performed without human control.
11. 11. The method according to any one of claims 1 to 10, wherein the second model (204) is configured to drive a technical system (112) depending on the prediction (212), and the technical system (112) is driven and controlled depending on the prediction (212) of the second model (204).
12. An apparatus (100) for generating an anomaly, comprising: The device (100) comprises: At least one processor (102); At least one memory (104); Including, The at least one processor (102) is configured to execute instructions stored in the at least one memory (104) that, when executed by the at least one processor (102), perform a method according to any one of claims 1 to 11.
1. An apparatus (100).
13. A computer program comprising:
12. The method according to claim 1, further comprising: instructions executable by at least one processor (102), the instructions, when executed by the at least one processor (102), performing the method according to any one of claims 1 to 11. A computer program comprising: