DE102022207266A1 - Apparatus and method for augmenting an image for self-supervised learning - Google Patents

Abstract

A computer-implemented image augmentation method for self-supervised learning, comprising the following steps: receiving (S21) at least one image (I), a plurality of augmentation groups ({gi}), and a plurality of hyperparameters ({Pgi}) that characterize the groups. Initialize (S22) an empty list of augmentation sequences (A). Repeating (S23) the steps of sampling a group g according to the first hyperparameter ({Pgi}) and sampling, according to the second hyperparameter (Ng), augmentations from the group (g) and appending the augmentations to the list (A ). Then the sampled augmentation sequences are applied to the image according to the list for image (I).

Description

State of the art

Self-supervised learning (SSL) is a machine learning technique. It learns from unlabeled sample data. It can be viewed as an intermediate form between supervised and unsupervised learning. It is based on an artificial neural network. The neural network learns in two steps. First, the task is solved based on pseudolabels, which helps initialize the network weights. Second, the actual task is performed using supervised or unsupervised learning.

SSL has experienced an explosion of interest in recent years, with great progress using SSL as a pre-training method for classification. A wide variety of SSL techniques have been developed, such as using different objective functions and optimization strategies, with continually improving performance on benchmarks such as Imagenet.

The data augmentation procedures in self-supervised learning (SSL) were predominantly set manually. For supervised learning, there are many methods for learning the right data augmentations, but not for SSL.

Therefore, it is proposed to optimize the data augmentation hyperparameters of SSL algorithms.

Advantages of the invention

The invention proposes, for. B. Use Bayesian Neural Pipeline Search (NEPS) optimization to search for optimal configurations. This presents an improved customized SSL-specific augmentation policy. The learned configurations improve linear classification performance in all settings we evaluated. In other words, a specific parameterization of the augmentation space with strong performance. This results in better performance and less manual labeling work.

Disclosure of the invention

In a first aspect, the invention relates to a method for image augmentation for self-supervised learning according to claim 1. The motivation behind self-supervised learning is to first learn general feature representations using unlabeled data and then to fine-tune these representations on downstream tasks using a few labels . The question arises as to how to learn these useful representations without knowing the labels.

Self-supervised learning (SSL) is a machine learning paradigm in which a model, when fed unstructured data as input, automatically generates data labels that can be further used as ground truths in subsequent iterations. Optionally, the model uses the high confidence data labels among those generated to train the model using backpropagation in the next iterations like any other supervised learning model. The only difference is that the data labels used as ground truths are changed in each iteration.

It is proposed that the hyperparameters be optimized using Bayesian optimization. Bayesian optimization optimizes the parameterization of the hyperparameters e.g. B. by sampling them according to a detection function and testing how well the selected hyperparameters perform in a learning task. These steps are repeated several times until an optimal parameterization is found.

It is further proposed that an image classifier be trained with the augmented image of the first aspect of the invention.

• - Empfangen eines Sensorsignals, das Daten von einem Sensor umfasst,
• - Bestimmen eines Eingangssignals, das von dem Sensorsignal abhängt, und
• - Einspeisen des Eingangssignals in den Klassifizierer, um ein Ausgangssignal zu erhalten, das eine Klassifizierung des Eingangssignals charakterisiert.

In further aspects of the invention, it is envisaged to use the classifier trained with the above methods in a method comprising the following steps:

• - receiving a sensor signal that includes data from a sensor,
• - Determining an input signal that depends on the sensor signal, and
• - Feeding the input signal into the classifier to obtain an output signal that characterizes a classification of the input signal.

The invention can be used to analyze data obtained from a sensor. The sensor can determine measurements of the environment in the form of sensor signals, e.g. B. through digital images, e.g. B. video, radar, LiDAR, ultrasound, motion, thermal images can be given.

The classifier, e.g. B. a neural network can be equipped with such a structure that can be trained, e.g. B. identify and distinguish pedestrians and/or vehicles and/or street signs and/or traffic lights and/or road surfaces and/or human faces and/or medical anomalies in images from an imaging sensor. Alternatively, the classifier, e.g. B. a neural network, be equipped with such a structure that can be trained to identify spoken commands in signals from an audio sensor.

• - Bestimmen des Aktorsteuersignals in Abhängigkeit vom Ausgangssignal Vorzugsweise steuert der Aktor einen zumindest teilweise autonomen Roboter und/oder eine Herstellungsmaschine und/oder ein Zugangskontrollsystem.

Such classifiers can then be used to provide an actuator control signal for controlling an actuator, comprising all of the steps of the above method and further comprising the following step:

• - Determining the actuator control signal as a function of the output signal Preferably, the actuator controls an at least partially autonomous robot and/or a manufacturing machine and/or an access control system.

In a further aspect of the invention, a computer-implemented method for using the classifier to classify sensor signals is proposed. The classifier is adopted with the method according to any of the above aspects of the inventions, comprising the following steps: receiving a sensor signal comprising data from the imaging sensor, determining an input signal dependent on the sensor signal, and feeding the input signal to the classifier to obtain an output signal that characterizes a classification of the input signal.

In a further aspect of the invention, a computer-implemented method for using the classifier trained with the method according to any of the preceding aspects to provide an actuator control signal for controlling an actuator. Determining an actuator control signal depending on an output signal of the classifier, which may be determined as described by the previous section. It is proposed that the actuator controls an at least partially autonomous robot and/or a manufacturing machine and/or an access control system.

In a further aspect of the invention, a control system for operating the actuator is proposed. The control system includes the classifier adopted according to any of the foregoing aspects of the invention and is configured to operate the actuator according to an output of the classifier.

• 1 ein schematisches Flussdiagramm einer Ausführungsform der Erfindung;
• 2 ein Steuersystem, das einen zumindest teilweise autonomen Roboter steuert;
• 3 ein Steuersystem, das eine Herstellungsmaschine steuert;
• 4 ein Steuersystem, das einen Zugangskontrollsystem steuert;
• 5 ein Steuersystem, das ein Überwachungssystem steuert;
• 6 ein Steuersystem, das einen automatisierten persönlichen Assistenten steuert;
• 7 ein Steuersystem, das ein Bildgebungssystem steuert;
• 8 ein Trainingssystem zum Steuern des Klassifizierers.

Embodiments of the invention will be discussed in more detail with reference to the following figures. The figures show:

• 1 a schematic flow diagram of an embodiment of the invention;
• 2 a control system that controls an at least partially autonomous robot;
• 3 a control system that controls a manufacturing machine;
• 4 a control system that controls an access control system;
• 5 a control system that controls a monitoring system;
• 6 a control system that controls an automated personal assistant;
• 7 a control system that controls an imaging system;
• 8th a training system to control the classifier.

Description of the embodiments

In supervised learning, data augmentation algorithms typically outperform strategies with manually selected augmentation. Three algorithms that rely on randomly sampling augmentations from a fixed set of augmentations are known as TrivialAugment and SmartSamplingAugment. Other algorithms, such as AutoAugment, RandAugment, and SmartAugment, perform gradient-free search in the space of augmentation policies. It is also possible to apply gradient-based optimization to meta-learn pre-training hyperparameters for ECG data. In Self-Augment, the authors rely on the correlation between supervised and self-supervised evaluation performance to incorporate a rotation task to efficiently generate augmentation policies. SelfAugment is qualitatively different from previous approaches in the sense that it does not optimize for general downstream performance, but rather for rotational task performance. Additionally, it will likely not include rotations in the selected augmentation policy.

In contrast to these, our GroupAugment approach covers a more diverse space of augmentation strategies than existing methods and can e.g. B. Design augmentation strategies that are similar to manually designed SSL augmentation strategies. It generalizes existing approaches by optimizing group-specific sampling probabilities, the number of group-specific augmentations, and the total number of applied augmentations, while imposing no restrictions on the set of augmentations such as SelfAugment.

To optimize over the search spaces listed above, Bayesian optimization (BO) is used ( Mockus, J., Tiesis, V. and Zilinskas, A. “The application of Bayesian methods for seeking the extremum. Towards Global Optimization,” 2(117-129) , 1978.) with expert prior knowledge, see e.g. B. Hvarfner, C., Stoll, D., Souza, A., Lindauer, M., Hutter, F.: “Augmenting acquisition functions with user beliefs for Bayesian optimization”. In Proceedings of the International Conference on Learning Representations (ICLR'21), 2021.

This section introduces GroupAugment, an automated data augmentation algorithm that works on groups of augmentations (such as color or quality transformations) and designs sampling strategies across these groups. Furthermore, an empirical study is presented in which it is found that, unlike the data augmentation algorithms designed for supervised learning, GroupAugment robustly outperforms the baseline in all settings analyzed.

Given some groups of data augmentations (g _i ), GroupAugment uses a global hyperparameter and two sets of group-specific hyperparameters to generate a list of augmentation sequences that are sequentially applied to an image. The global hyperparameter 7 determines the number of augmentation sequences in this list. The group-specific hyperparameter P _gi determines the probability that the augmentation group _gi is chosen to produce the following augmentation sequence. N _gi determines the number of augmentations to sample uniformly without replacement to form an augmentation sequence for the augmentation group _gi . Pseudocode for a GroupAugment policy is provided in Algorithm 1:

Applying sampled augmentation sequences A to I

GroupAugment generalizes this idea to any set of augmentation groups and searches over more general spaces of sampling strategies. In one embodiment, GroupAugment is instantiated with five groups: color, geometric, non-rigid, quality, and exotic. See table for an example of specific augmentations and their grouping that were used: group augmentation Color ColorJitter(Brightness=0.4, Contrast=0.4, Saturation=0.4, Hue=0.1) ToGray() Solarize() Equalize() ChannelShuffle() geometric ShiftScaleRotate(Interpolation=cv2. I NTE R CUBIC) Horizontal Flip() Unrigid ElasticTransform(Alpha=0.5, Sigma=10, Alpha affine=5, Interpolation=cv2.INTER CUBIC) GridDistortion(Interpolation=cv2.INTER CUBIC) OpticalDistortion(Distortion limit=0.5, Displacement limit=0.5, Interpolation=cv2.INTER CUBIC) Quality GaussianBlur() GaussNoise() exotic RandomGridShuffle() Cutout(number of holes=4)

See the following table for an example description of GroupAugment's search space: Hyperparameters Type Area Log Prior default p color transformations Floating point number [0,1] Incorrect 0.5 p geometric transformations Floating point number [0,1] Incorrect 0.5 p non-rigid transform floating point [0,1] Incorrect 0.0 mations number p quality transformations Floating point number [0,1] Incorrect 0.0 p exotic transformations Floating point number [0,1] Incorrect 0.0 Number of color transformations Integer [0.5] Incorrect 1 Number of geometric transformations Integer [0.5] Incorrect 1 Number of non-rigid transformations Integer [0.5] Incorrect 1 Number of quality transformations Integer [0.5] Incorrect 1 Number of total group samples Integer [0.5] Incorrect 1

To optimize over the resulting search space, the same expert Bayesian optimization (BO) approach is used as above. Furthermore, since BO samples each group probability individually, the sampled probabilities are normalized, and therefore they do not necessarily add up to 1.

1 shows an exemplary image augmentation method for self-supervised learning.

The method starts with receiving (S21) at least one image (I), a plurality of augmentation groups ({g _i }), a first hyperparameter ({P _gi }), which characterizes group-specific sampling probabilities, a second hyperparameter ({N _gi }), which characterizes a number of group-specific augmentations, and a third hyperparameter (7), which characterizes a number of total group samples.

A step of initializing (S22) an empty list of augmentation sequences (A) is then carried out.

A loop (S23) is then executed in which a counter is increased step by step until the value of the third hyperparameter (T) is reached. The loop includes a step of sampling a group g according to the first hyperparameter ({P _gi }) and sampling, according to the second hyperparameter (N _g ), augmentations from the sampled group (g). The sampled augmentation is added to list (A).

After multiple augmentations have been sampled and the loop is completed, the sampled augmentations are applied to an image according to the defined sequence of the list (S25).

Optionally, an image classifier, preferably a neural network, is trained according to self-supervised learning with the augmented image from step S25.

In 2 an embodiment of an actuator with a control system 40 is shown. The actuator and its environment are collectively called an actuator system. A sensor 30 detects a state of the actuator system at preferably uniformly spaced distances. The sensor 30 may include multiple sensors. Preferably, the sensor 30 is an optical sensor that records images of the environment. An output signal S of the sensor 30 (or in the case that the sensor 30 includes a plurality of sensors, an output signal S for each of the sensors) that encodes the detected condition is transmitted to the control system 40.

Thereby, the control system 40 receives a stream of sensor signals S. It then calculates a series of actuator control commands A depending on the stream of sensor signals S, which are then transmitted to an actuator unit 10 which converts the control commands A into mechanical movements or changes physical quantities. For example, the actuator unit 10 can convert the control command A into an electrical, hydraulic, pneumatic, thermal, magnetic and/or mechanical movement or change. Specific but non-limiting examples include electric motors, electroactive polymers, hydraulic cylinders, piezoelectric actuators, pneumatic actuators, servomechanisms, solenoids, stepper motors, etc.

The control system 40 receives the stream of sensor signals S from the sensor 30 in an optional receiving unit 50. The receiving unit 50 transforms the sensor signals S into input signals x. Alternatively, in a case without a receiving unit 50, each sensor signal S can be directly accepted as an input signal x. The input signal x can be given, for example, as an extract from the sensor signal S. Alternatively, the sensor signal S can be processed to yield the input signal x. The input signal x includes image data corresponding to an image recorded by the sensor 30. In other words, the input signal x is provided according to the sensor signal S.

The input signal x is then passed on to the image classifier 60, which can be given, for example, by an artificial neural network and with augmented images created by the method of 1 were obtained, trained.

The classifier 60 is parameterized by parameters φ, which are stored in and provided by the parameter storage St ₁ .

The classifier 60 determines output signals y from the input signals x. The output signal y includes information that assigns one or more labels to the input signal x. The output signals y are transmitted to an optional conversion unit 80, which converts the output signals y into the control commands A. The actuator control commands A are then transmitted to the actuator unit 10 to control the actuator unit 10 accordingly. Alternatively, the output signals y can be accepted directly as the control commands A.

The actuator unit 10 receives the actuator control commands A, is controlled accordingly and carries out an action in accordance with the actuator control commands A. The actuator unit 10 may include control logic that transforms the actuator control command A into a further control command, which is then used to control the actuator 10.

In further embodiments, the control system 40 may include the sensor 30. In still further embodiments, the control system 40 may alternatively or additionally include the actuator 10.

In one embodiment, the classifier 60 may be configured to identify lanes on a road ahead, e.g. B. by classifying a road surface and markings on the road and identifying tracks as areas of the road surface between the markings. Based on an output from a navigation system, an appropriate lane for following a selected path may then be selected, and depending on a current lane and the target lane, a decision may then be made as to whether the vehicle 60 should change lanes or remain in the current lane. The control command A can then be calculated, e.g. B. by retrieving a predefined movement pattern from a database corresponding to the identified action.

Likewise, after identifying street signs or traffic lights, depending on an identified type of street sign or an identified state of the traffic lights, corresponding restrictions can then be placed on possible movement patterns of the vehicle 60, e.g. B. retrieved from a database, a future path of the vehicle 60 may be calculated according to the constraints, and the actuator control command A may be calculated to steer the vehicle to execute the trajectory.

Likewise, after identifying pedestrians and/or vehicles, projected future behavior of the pedestrians and/or vehicles may be estimated, and based on the estimated future behavior, a trajectory may then be selected to avoid a collision with the pedestrian and/or vehicle Vehicle to avoid, and the actuator control command A can be calculated to steer the vehicle to execute the trajectory.

In still further embodiments, it can be imagined that the control system 40 controls a display 10a of an actuator 10, the display 10a displaying the control command or the like can.
In other embodiments, the display 10a may be an output interface to a display device, such as a display, a light source, a speaker, a vibration motor, etc., that may be used to generate a sensory-perceivable output signal based on the feature representation can be extracted by the video feature extractor. The sensory perceptible output signal can directly indicate the extracted features, but can also represent a derived sensory perceptible output signal, e.g. B. for use in guidance, navigation or other type of control of a computer controlled system.

Further, the control system 40 may include a processor 45 (or a plurality of processors) and at least one machine-readable storage medium 46 storing instructions that, when executed, cause the control system 40 to carry out a method according to an aspect of the invention.

In a preferred embodiment of 2 the control system 40 is used to control the actuator, which is an at least partially autonomous robot, e.g. B. an at least partially autonomous vehicle 100.

The sensor 30 may include one or more video sensors and/or one or more radar sensors and/or one or more ultrasonic sensors and/or one or more LiDAR sensors and/or one or more position sensors (such as GPS). Some or all of these sensors are preferably, but not necessarily, integrated into the vehicle 100. Alternatively or additionally, the sensor 30 may include an information system for determining a state of the actuator system. An example of such an information system is a weather information system that determines a current or future state of the weather in the environment 20.

For example, the classifier 60 can detect, for example, objects in the vicinity of the at least partially autonomous robot using the input signal x. The output signal y may include information that characterizes where objects are located in the vicinity of the at least partially autonomous robot. The control command A can then be determined according to this information, for example to avoid collisions with the detected objects.

The actuator unit 10, which is preferably integrated in the vehicle 100, can be provided by a brake, a drive system, a motor, a drive train or a steering system of the vehicle 100. The actuator control commands A can be determined so that the actuator (or actuators) unit 10 is controlled so that the vehicle 100 avoids collisions with the detected objects. The detected objects can also be classified according to what the classifier 60 considers most likely, e.g. B. pedestrians or trees, and the actuator control commands A can be determined depending on the classification.

In further embodiments, the at least partially autonomous robot may be another mobile robot (not shown) that can move, for example, by flying, swimming, diving or walking. The mobile robot can be, among other things, an at least partially autonomous lawn mower or an at least partially autonomous cleaning robot. In all of the above embodiments, the actuator control command A may be determined so that a drive unit and/or a steering and/or a brake of the mobile robot are controlled so that the mobile robot can avoid collisions with the identified objects.

In a further embodiment, the at least partially autonomous robot may be a garden robot (not shown) that uses the sensor 30, preferably an optical sensor, to determine a condition of plants in the environment 20. The actuator unit 10 can be a nozzle for spraying chemicals. Depending on an identified species and/or condition of the plants, an actuator control command A may be determined to cause the actuator unit 10 to spray the plants with an appropriate amount of appropriate chemicals.

In still further embodiments, the at least partially autonomous robot may be provided by a household appliance (not shown), such as. B. a washing machine, a stove, an oven, a microwave or a dishwasher. The sensor 30, e.g. B. an optical sensor can detect a state of an object that is to be subjected to processing by the household appliance. For example, in the case that the household appliance is a washing machine, the sensor 30 can indicate a state of the Detecting laundry in the washing machine. The actuator control signal A can then be determined depending on a detected material of the laundry.

In 3 An embodiment is shown in which a control system 40 is used to control a manufacturing machine 11, e.g. B. a punching device, a cutting tool or a deep hole drill, a manufacturing system 200 z. B. to control as part of a production line.

The control system 40 controls an actuator unit 10, which in turn controls the manufacturing machine 11.

The sensor 30 can be an optical sensor which has properties such as: B. a manufactured product 12 recorded. The classifier 60 can determine a condition of the manufactured product 12 from these detected properties. The actuator unit 10, which controls the manufacturing machine 11, can then be controlled for a subsequent manufacturing step of the manufactured product 12 depending on the specific state of the manufactured product 12. Or it can be imagined that the actuator unit 10 is controlled during the production of a subsequent manufactured product 12 depending on the specific state of the manufactured product 12.

In 4 An embodiment is shown in which the control system controls an access control system 300. The access control system can be designed to physically control access. It may include a door 401, for example. The sensor 30 is designed to detect a scene that is relevant to deciding whether or not to grant access. For example, it can be an optical sensor to provide image or video data for detecting a person's face. The classifier 60 may be designed to interpret this image or video data, e.g. B. by matching identities with known people stored in a database, thereby determining an identity of the person. The actuator control signal A can then, depending on the interpretation of the classifier 60, e.g. B. be determined according to the specific identity. The actuator unit 10 can be a lock that grants access or not depending on the actuator control signal A. Non-physical, logical access control is also possible.

In 5 An embodiment is shown in which the control system 40 controls a monitoring system 400. This embodiment is largely similar to that in 5 Embodiment shown identical. Therefore, only the different aspects are described in detail. The sensor 30 is designed to detect a monitored scene. The control system does not necessarily control an actuator 10, but rather a display 10a. For example, the machine learning system 60 may determine a classification of a scene, e.g. B. whether the scene detected by the optical sensor 30 is suspicious. The actuator control signal A, which is transmitted to the display 10a, can then z. B. be designed to cause the display 10a to adjust the displayed content depending on the particular classification, e.g. B. highlight an object that is considered suspicious by the machine learning system 60.

In 6 An embodiment is shown in which the control system 40 is used to control an automated personal assistant 250. The sensor 30 can be an optical sensor, for example. B. to receive video images of gestures of a user 249. Alternatively, the sensor 30 can also be an audio sensor, for example. B. to receive a voice command from the user 249.

The control system 40 then determines the actuator control commands A for controlling the automated personal assistant 250. The actuator control commands A are determined according to the sensor signal S of the sensor 30. The sensor signal S is transmitted to the control system 40. For example, the classifier 60 may be designed to, e.g. B. execute a gesture recognition algorithm to identify a gesture made by the user 249. The control system 40 may then determine an actuator control command A for transmission to the automated personal assistant 250. It then transmits the actuator control command A to the automated personal assistant 250.

For example, the actuator control command A may be determined according to the identified user gesture recognized by the classifier 60. It may then include information that causes the automated personal assistant 250 to retrieve information from a database and output that retrieved information in a form suitable for receipt by the user 249.

In further embodiments, it can be imagined that instead of the automated personal assistant 250, the control system 40 controls a household appliance (not shown) that operates according to the identified user gesture. The household appliance can be a washing machine, a stove, an oven, a microwave or a dishwasher.

In 7 1 is an embodiment of a control system 40 for controlling an imaging system 500, for example an MRI device, an X-ray imaging device or an ultrasound imaging device. The sensor 30 may be, for example, an imaging sensor. The machine learning system 60 can then determine a classification of all or part of the captured image. The actuator control signal A can then be selected according to this classification, thereby controlling the display 10a. For example, the machine learning system 60 may interpret an area of the captured image as potentially anomalous. In this case, the actuator control signal A may be determined to cause the display 10a to display the image and highlight the potentially abnormal area.

In 8th an embodiment of a training system 500 is shown. The training device 500 includes a provision system 51 that provides input images from a training data set. The input images are fed into the neural network 52 to be trained, which determines output variables from them. The output variables and the input images are supplied to an evaluator 53 which determines acute hyper/parameters therefrom which are transferred to the parameter memory P in which they replace the current parameters.

The procedures performed by the training device 500 may be implemented as a computer program stored on a machine-readable storage medium 54 and executed by a processor 55.

The term “computer” covers any device for processing predefined calculation instructions. These calculation instructions can be in the form of software or in the form of hardware or even in a mixed form of software and hardware.

It is further understood that the procedures cannot only be implemented entirely in software, as described. They can also be implemented in hardware or in a mixed form of software and hardware.

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent literature cited

• Mockus, J., Tiesis, V. and Zilinskas, A. “The application of Bayesian methods for seeking the extremum. Towards Global Optimization”, 2(117-129) [0020]

Claims (10)

Computer-implemented method (20) for image augmentation for self-supervised learning, comprising the following steps: receiving (S21) at least one image (I), a plurality of augmentation groups ({g _i }), a first hyperparameter ({P _gi }), which characterizes group-specific sampling probabilities, a second hyperparameter ({N _gi }) which characterizes a number of group-specific augmentations, and a third hyperparameter (T) which characterizes a number of total group samples; Initialize (S22) an empty list of augmentation sequences (A): Repeat (S23) until a counter reaches the value of the third hyperparameter (T): Sample a group g according to the first hyperparameter ({{P _gi }); Sampling, according to the second hyperparameter (N _g ), augmentations from the group (g); Append the augmentations to the list (A); Apply (S24) the sampled augmentation sequences according to the list to the image (l). Procedure according to Claim 1 , where the step of sampling according to the second hyperparameter N _g is carried out by uniform sampling without replacement. Method according to one of the preceding claims, wherein the hyperparameters are optimized using Bayesian optimization. Method according to one of the preceding claims, wherein an image classifier (60) is trained by self-supervised learning with the augmented image. Procedure according to Claim 4 , wherein the classifier (60) is trained to classify sensor signals, the classifier (60) comprising the following steps: - receiving a sensor signal (S) comprising data from a sensor (30), - determining an input signal (x) , which depends on the sensor signal (S), and feeding the input signal (x) into the classifier (60) to obtain an output signal (y) which characterizes a classification of the input signal (x). Computer-implemented method for using a classifier (60) that uses the method according to Claim 5 is trained to provide an actuator control signal (A) for controlling an actuator (10) that carries out all steps of the method Claim 5 and further comprises the following step: - determining the actuator control signal (A) as a function of the output signal (y). Procedure according to Claim 6 , wherein the actuator (10) controls an at least partially autonomous robot (100) and/or a manufacturing machine (200) and/or an access control system (300). Computer program designed to cause a computer to carry out the method according to one of the Claims 1 until 7 with all of its steps if the computer program is executed by a processor. Machine-readable storage medium on which the computer program is written Claim 8 is stored. System designed to carry out the method according to one of the Claims 1 until 7 to carry out.

Images