WO2020030722A1

WO2020030722A1 - Sensor system including artificial neural network configured to perform a confidence measure-based classification or regression task

Info

Publication number: WO2020030722A1
Application number: PCT/EP2019/071277
Authority: WO
Inventors: Steve DIAS DA CRUZ; Hans-Peter Beise; Udo Schröder
Original assignee: Iee International Electronics & Engineering S.A.
Priority date: 2018-08-09
Filing date: 2019-08-08
Publication date: 2020-02-13
Also published as: LU100902B1

Abstract

A sensor system includes at least one sensor and an evaluation device that is configured for receiving input data x from the at least one sensor, and that comprises at least one trained artificial neural network that is configured to perform a classification or regression task on input data x, wherein the at least one trained artificial neural network comprises an M_Θ-module as an implementation of a machine learning based method for the classification or regression task with trainable parameters Θ. The sensor system further comprises a confidence measure module arrangement that includes a D_p- module as implementation of a machine learning based method that is configured to learn a representation of the training dataset with trainable parameters, and an E-module as implementation of a measure to determine how far the input data x are from the training dataset using the information of DΦ. The confidence measure module arrangement is configured to utilize said output being combined with the output of M_Θ to decide whether the module is allowed to perform an action (classification or regression task based on the input data x).

Description

Sensor system including artificial neural network configured to perform a confidence measure-based classification or regression task

Technical field

[0001] The invention relates to a sensor system including at least one trained artificial neural network to perform, for example, a classification or regression task on input data received from the sensor or the sensors. The at least one trained artificial neural network comprises a module as an implementation of a machine learning based method for the classification or regression task. The invention further relates to a method of operating such sensor system.

Background of the Invention

[0002] The decision process of machine learning based algorithms, and especially artificial neural networks (in the following also referred to as neural network for briefness), can usually not be influenced once the algorithm has been implemented and a device in which it has been installed is in the field. Consequently, particularly for safety-critical sensor applications, there is no measure to decide on whether an algorithm should be allowed to take an action for a new input. That is, there is no inherent measure that checks whether the model (i.e. the neural network) is trained in a way that it generalizes as expected to some new input of the device. Usually, the algorithm is forced to take an action for every input, although the algorithm and model has no possibility to detect whether the input is exotic, or simply whether the input is in some sense far from what it has seen during training. Since input data can be very different from the initial training dataset on which the model was trained on, the model does not know how to handle such input correctly and it will, for example, classify the input in most cases wrongly.

[0003] It would be desirable in connection with a sensor system including at least one sensor and at least one trained artificial neural network to perform a classification or regression task for an input to the sensor system to have a possibility to check during the classification or regression task whether the neural network is allowed to perform an action or whether the neural network should ask for human interaction or warn the sensor system or a user that the sensor system is not able make a meaningful decision.

[0004] In the article by Juutilainen, I. et al.,“A Method for Measuring Distance From a Training Data Set’, Communications in Statistics-Theory and Methods, 36: 2625-2639, 2007 (DOI: 10.1080/03610920701271129), a method is proposed for measuring the distance between a training data set and a single, new observation. The novel distance measure reflects the expected squared prediction error when a quantitative response variable is predicted on the basis of the training data set using the distance weighted k-nearest-neighbor method. It is further presented that the distance measure correlates well with the true expected squared prediction error in practice. The distance measure can be applied, for example, in assessing the uncertainty of prediction.

[0005] As the k-nearest-neighbors are involved, this approach works for very simple datasets only, i.e. not for images, because it would require to save all the training data of a training data set on a system, or to have a connection to a server. Classifying a new image would require to compare the new image to each single image in the training data set, which is obviously too slow for a large number of applications. Furthermore, distances between images using the pixel values cannot be considered meaningful measures.

Object of the invention

[0006] It is therefore an object underlying the present invention to provide a sensor system of the kind considered without at least some of the above described shortcomings.

[0007] This object is achieved by a system comprising the features of claim 1. The object is further achieved by a method of operating a sensor system comprising the features of claim 12.

General Description of the Invention

[0008] The invention provides a sensor system including at least one sensor and an evaluation device that is configured for receiving input data from the at least one sensor, and that comprises at least one trained (artificial) neural network that is configured to perform a classification or regression task on input data received from the at least one sensor. The at least one trained neural network comprises an M_g -module as an implementation of a machine learning based method for the classification or regression task with trainable parameters Q. The system further comprises a confidence measure module arrangement used in combination with the at least one trained neural network to decide when to perform a decision or regression on the input data x of the at least one sensor and when not. The confidence measure module arrangement includes:

a ΰf -module as implementation of a machine learning based method (designed as a neural network) that is configured to learn a representation of the training dataset with trainable parameters f, wherein the representation of the training dataset is lower-dimensional than the training dataset, and an E-module as implementation of a measure to determine how far the input data x are from the training dataset using the information

wherein the ΰf -module is configured to provide for

(x)-output of the neural network which learned the representation of the training dataset, said output being used by the E-module to determine how different the input data are from what has been seen during training, the E-module is configured to provide for an E(Ef (c),c)- output of the neural network which determines how far the input data x are from the training dataset using (x), said output being combined with M_g to decide whether the neural network is allowed to perform an action (classification or regression task), and the M_g -module is configured to provide for an M_g (x)-output of a classification or regression model using the input data (x), and the confidence measure module arrangement is configured to utilize said E(Ef (c),c)- output and said M_g (x)-output in combination to decide to perform an action, upon which the output will be the classification or regression based on the input data x. Further, the evaluation device is configured to generate an output signal representing a result of the classification or regression task. [0009] The phrase “being configured to”, as used in this application, shall in particular be understood as being specifically programmed, laid out, furnished or arranged.

[0010] In other words, this invention proposes to use a confidence measure module arrangement on which a lower-dimensional representation of the training dataset is stored/encoded (e.g. by means of an own neural network). This confidence measure module arrangement is operated in combination with the neural network for performing classification or regression tasks in order to decide when the neural network should be allowed to perform a decision or regression task and when not. As a consequence, misclassifications can advantageously be avoided, and the reliability of safety-critical sensor system applications can be increased. The classification or regression algorithm is adapted to learn during training how to efficiently store the training dataset and how to use it.

[0011] The sensor system of the invention allows for the self-verification of the capability of handling the input data x to an evaluating device correctly during lifetime, wherein input data x received during lifetime are being compared with the data seen during training. To this end, the invention provides for training the neural network of the confidence measure module arrangement, which neural network is configured to learn to represent the training data in a compact and useful way. This neural network can be used during the classification and regression task to determine if the module has the capability to use and process the input correctly. The neural network can also be a part of the module that is trained for the actual classification/regression task.

[0012] The various modules of the confidence measure module arrangement may comprise at least one neural network.

[0013] The proposed sensor system that employs a neural network for evaluation has the advantage of requiring comparatively less hardware. Further, there is no need for storing the complete training dataset but only the parameters of the

module. For classifying a new sensor signal it can be omitted to compare the new sensor signal with each signal of the training dataset. Besides the lower required storage space, this also has the advantage of a faster computation, allowing to use the sensor system for real-time applications.

[0014] Preferably, the at least sensor is formed as a radar sensor or as an optical camera. The term“optical camera”, as used in this application, shall in particular encompass, without being limited to, an optical digital single-shot and a video camera for visible light. The last mentioned embodiments of the proposed sensor system can beneficially be employed, without being limited to, in automotive vehicle exterior applications for monitoring and surveying of other tracking participants or in automotive vehicle interior applications for passenger cabin surveillance such as detection of left-behind pets and/or children, vital sign monitoring, vehicle seat occupancy detection and anti-theft alarm.

[0015] In preferred embodiments of the sensor system, the machine learning based method implementation

comprises a variational autoencoder (VAE). VAEs are known to be efficiently employable for dimensionality reduction in a generative approach to classification.

[0016] For sensor system embodiments of the latter type, the E-module is preferably configured to calculate an /₂-error, l₂ between an

output of the autoencoder, denoted by Y, and the input data x as the measure to detect whether the input data is close to the training dataset (module E(ΰf (c), c)). In this way, the closeness of the input data x to the training data set can readily be assessed by comparing a single number to a predefined threshold.

[0017] In preferred embodiments of the sensor system, wherein in case that a value of the measure describing how far the input data x are from the training dataset is lower than a predefined threshold value, the M^-module is configured to perform a classification or regression task based on the input data x. Otherwise, the Mg-module is configured to generate a signal indicating that it is hindered to perform such action. In this way, an output signal ambiguity of the sensor system can be avoided.

[0018] Preferably, the D^-module and the M^-module are trained together/or in parallel and are configured to interact with each other. In this way, an efficiency of the sensor system can be improved. Further preferably, the Mg-module is configured to use the input data x as well as the lower-dimensional representation Z of the input data x in order to perform a classification or regression task, by which misclassifications can further be reduced, and the reliability of safety-critical sensor system applications can be increased.

[0019] Particularly for sensor system applications with a reduced safety criticality, the D^-module and the M^-module are preferably trained independently from one another and are combined in the system after training, by which a training effort can be reduced.

[0020] In preferred embodiments, in which the sensor system comprises at least one optical camera, the

module comprises a region of interest algorithm that is configured for proposing only predefined interesting regions in an image. This can enable a classification regarding only the predefined interesting regions by the M_g- module and can allow for a faster performance of the sensor system. In this case, the Mg-module is preferably configured to perform a classification for each of the predefined interesting regions.

[0021] In such embodiments of the sensor system, in which the

module comprises a region of interest algorithm, the region of interest algorithm is preferably configured to be background-independent, by which a performance of the sensor system can be expedited.

[0022] In another aspect of the present invention, a method of operating a sensor system is provided. The sensor system includes at least one sensor and an evaluation device that is configured for receiving input data x from the at least one sensor. The evaluation device comprises at least one trained artificial neural network with trained parameters Q that is configured to perform a classification or regression task on data received from the at least one sensor. The method comprises at least the following steps:

providing input data x of the at least one sensor to the evaluation device, applying a machine learning based method implementation

that has learned a representation of a training dataset with trainable parameters f, wherein the representation of the training dataset is lower-dimensional than the training dataset, to the input data of the at least one sensor for providing a ΰ_f (x)-output, by using the information of ΰ_f (x), applying a measure implementation E for determining how far the input data x are from the training dataset and obtaining a measure output E(0_y(c),c), comparing the obtained measure output E(r_f (c), c) with at least one predefined threshold value for the measure,

performing a classification or regression task on the input data x received from the at least one sensor, if a result of the comparison fulfills a predefined condition, and

generating an output signal representing a result of the classification or regression, if a classification or regression has been performed.

The aforementioned benefits described in context with the sensor system in accordance with the invention also apply to the proposed method.

[0023] Preferably, the step of applying a measure implementation E for determining how far the input data x are from the training dataset includes calculating an Z₂-error as a measure according to l₂ = In this

way, an appropriate measure can readily be provided.

[0024] In preferred embodiments, the method further comprises the step that a signal is generated that indicates that the step of performing a classification or regression on the input data x has been hindered, if the result of the comparison fails to fulfill the predefined condition. In this way, an output signal ambiguity of the sensor system can be avoided.

[0025] There exists prior art taking into account a confidence measure to determine how likely the algorithm can take a decision. Examples of such prior art is disclosed in the US20110087627, US20170185893, US5052043, and US5912986. [0026] US patent application US20110087627 discloses a system and method for generating a prediction using neural networks, training a plurality of neural networks with training data, calculating an output value for each of the plurality of neural networks based at least in part on input evaluation points, applying a weight to each output value based at least in part on a confidence value for each of the plurality of neural networks; and generating an output result.

[0027] US patent application US20170185893 discloses a computer-implemented method of incrementally training a confidence assessment module that calculates a confidence value indicative of the extent to which a code associated with a patient's encounter with a healthcare organization is proper. The method comprises assessing, with the confidence assessment module, a training corpus comprised of a plurality of coded encounters, to produce resultant confidence values associated with each encounter; comparing the resultant confidence values to a target confidence value; and, adjusting variables within the confidence assessment module to produce resultant confidence values closer to the target confidence value.

[0028] US patent US5052043 discloses a method, for a neural network, which through controlling back propagation and adjustment of neural weight and bias values through an output confidence measure rapidly and accurately adapts its response to actual changing input data. The results of an appropriate actual unknown input are used to adaptively re-train the network during pattern recognition. By limiting the maximum value of the output confidence measure at which this re-training will occur, the network re-trains itself only when the input has changed by a sufficient margin from initial training data such that this re-training is likely to produce a subsequent noticeable increase in the recognition accuracy provided by the network.

[0029] US patent US5912986 discloses a method for use in a neural network- based optical character recognition system for accurately classifying each individual character extracted from a string of characters. Specifically, a confidence measure, associated with each output of, e.g., a neural classifier, is generated through use of all the neural activation output values. Each individual neural activation output provides information for a corresponding atomic hypothesis of an evidence function. This hypothesis is that a pattern belongs to a particular class. Each neural output is transformed through a predefined monotonic function into a degree of support in its associated evidence function. These degrees of support are then combined through an orthogonal sum to yield a single confidence measure associated with the specific classification then being produced by the neural classifier.

[0030] Therefore, this prior art does not interfere with the approach of the present invention defined by claims 1 and 12, respectively.

[0031] Advantageous developments of the invention are defined in the dependent claims.

Brief Description of the Drawings

[0032] Further details and advantages of the present invention will be apparent from the following detailed description of not limiting embodiments with reference to the attached drawing, wherein:

Fig. 1 schematically illustrates a sensor system in accordance with the invention installed in a vehicle in a side view,

Fig. 2 depicts a schematic of an embodiment of the confidence measure module arrangement of the sensor system pursuant to Fig. 1 , and

Fig. 3 depicts a schematic of a further embodiment of the confidence measure module arrangement of the sensor system pursuant to Fig. 1.

Description of Preferred Embodiments

[0033] Fig. 1 schematically illustrates a sensor system 10 in accordance with the invention installed in a vehicle 18, which is designed as a sedan passenger car, in a side view. The sensor system 10 is configured to be used as an automotive vehicle interior sensing system for a detection of for instance, but not limited to, left-behind pets and/or children, vital sign monitoring, vehicle seat occupancy detection for seat belt reminder (SBR) systems, and/or anti-theft alarm. To that end, the sensor system 10 includes a sensor 12 that is formed as an optical camera. The optical camera may be fixedly or movably connected to a chassis of the vehicle 18, or it may be integrated in the vehicle dashboard. [0034] The sensor system 10 further comprises an evaluation device 14 that is configured for receiving input data x from the optical camera. The evaluation device 14 includes a trained artificial neural network 16 (Fig. 2) that is configured to perform a classification or regression task on input data x received from the optical camera. The artificial neural network 16 includes an M_g -module as an implementation of a machine learning based method for the classification or regression task with trainable parameters Q.

[0035] The sensor system 10 further comprises a confidence measure module arrangement used in combination with the trained artificial neural network 16 to decide when to perform a decision or regression on the input data x of the optical camera and when not. Fig. 2 depicts a schematic of an embodiment of the confidence measure module arrangement of the sensor system 10 pursuant to Fig. 1. The confidence measure module arrangement comprises the following modules:

a ΰf -module as implementation of a machine learning based method (for instance artificial neural network) that is configured to learn a representation of the training dataset with trainable parameters f, wherein the representation of the training dataset is lower-dimensional than the training dataset, and

an E-module as implementation of a measure to determine how far the input data x are from the training dataset using the information of ΰ_f .

As shown in Fig. 2 the ΰ_f -module is configured to provide for an ΰ_f (x)-output of a neural network which learned the representation of the training dataset, said output being used by the E-module to determine how different the input data x are from what has been seen during training. Further, the E-module is configured to provide for an E(r_f (c), c) -output of a neural network which determines how far the input data x are from the training dataset using ΰ_f (x), said output being combined with M_g to decide whether the trained artificial neural network is allowed to perform an action (classification or regression). Still further, the M_g -module provides for an M_g (x)-output of a classification or regression model using the input data x, said output in combination with the output of the module and the output of E-module being adapted to decide to perform an action, upon which the output will be the classification or regression based on the input data x.

[0036] W_f and M_g should preferably be trained together/or parallel and can possibly interact with each other to improve the efficiency of the whole sensor system 10. However, both modules can also be trained independently and only be combined in the sensor system 10 after training.

[0037] Alternatively, a structure can be implemented in which ΰ_f and M_g are not trained together. For example, ΰ_f could be a“region of interest (ROI)” algorithm, proposing only the interesting regions in an image of the optical camera. This algorithm would be optimized to be background-independent. Then for each region, the module M_g, optimized for classifications, could perform a classification task.

[0038] Fig. 3 shows a schematic of a further embodiment of the confidence measure module arrangement of the sensor system pursuant to Fig. 1. For the data representation module ΰ_f , a variational autoencoder (VAE) may be used in the example which can learn to represent the training dataset in a lower- dimensional representation. The skilled person will appreciate that the variational autoencoder (VAE) is only one possible example of a neural network structure which can be used and that other autoencoder and/or neural network models could also be used as well.

[0039] The lower-dimensional representation of the input data x (in Fig. 3 also denoted by capital letter X) is denoted by Z in the following. In order to detect whether the input data are in some sense close to the training dataset (module (Ef (c),c)), one could calculate the /₂-error l₂ between the

output of the autoencoder, denoted by Y, and the input X. If the input X is close to the training dataset, then the lower-dimensional representation Z of X is meaningful and consequently the reconstruction by Y should be close to X (since this is the initial goal of the autoencoder). The term “close” in latter can be interpreted in a wide sense, for instance close in norm distance or statistical measures. The actual distance is determined by the design of the autoencoder. If the error between X and Y is lower than a predefined threshold value, than the module M_g is allowed to perform a classification or regression task based on X. Otherwise, the module is not allowed to do so and the module may inform the user or system that it cannot perform an action.

[0040] If ΰf and M_g are trained together, M_g can use the input X as well as the lower-dimensional representation Z of X in order to perform a classification or regression task. Further, the lower-dimensional representation could be optimized since M_g learns how to handle Z and X.

[0041] One can also imagine a structure in which ΰ_f and M_g are not trained together. For example, ΰ_f could be a “region of interest (ROI)” algorithm, proposing only the interesting regions in an image. This algorithm would be optimized to be background-independent. Then for each region, the module M_g, optimized for classifications, could perform a classification task.

[0042] In the following, an embodiment of a method of operating the sensor system 10 pursuant to Fig. 1 will be described with reference to Figs. 2 and 3. In preparation of operating the sensor system 10, it shall be understood that all involved units and devices are in an operational state and configured as illustrated in Fig. 1.

[0043] In a first step 30 of the method, input data x of the optical camera are provided to the evaluation device 14. The data representation module ϋ_f is applied to the input data x for providing a ΰ_f (x)-output in another step 32. Then, in a next step 34, by using the information of ΰ_f (x), the E-module is applied for determining how far the input data x are from the training dataset, expressed for instance by the Z₂-error, and for obtaining a measure output E(r_f (c), c).

[0044] The obtained measure output E(r_f (c), c) is compared in a next step 36 of the method with a predefined threshold value for the measure, i.e. the Z₂-error. If a result of the comparison fulfills a predefined condition, which in this specific embodiment is given by the Z₂-error being lower than the predefined threshold value, then a classification or regression task is performed on the input data x received from the optical camera in another step 38. If the result of the comparison fails to fulfill the predefined condition, a signal is generated in a further step 40 that indicates that the step 38 of performing a classification or regression task on the input data x has been hindered.

[0045] If a classification or regression task has been performed, an output signal 20 representing a result of the classification or regression task is generated in a final step 42.

LIST OF REFERENCE SYMBOLS

10 sensor system

12 sensor

14 evaluation device

16 artificial neural network

18 vehicle

20 output signal

x, X input data Y autoencoder output

z lower-dimensional representation of input data

Method steps:

30 provide input data of optical camera to evaluation device

32 apply data representation module ϋ_f to input data for providing ϋ_f (x)- output

34 apply module E to ϋ_f (x)-output for obtaining output E(ϋ_f (x),x) as measure for how far the input data are from training dataset

36 compare measure output with predefined threshold value

38 perform classification or regression task on input data

40 generate signal indicating that performing a classification or regression task has been hindered

42 generate output signal representing result of classification or regression task

Claims

1. A sensor system (10) including at least one sensor (12) and an evaluation device (14) that is configured for receiving input data (x) from the at least one sensor (12), and that comprises at least one trained artificial neural network (16) that is configured to perform a classification or regression task on input data (x) received from the at least one sensor (12), wherein the at least one trained artificial neural network (16) comprises an M_g -module as an implementation of a machine learning based method for the classification or regression task with trainable parameters Q, characterized in that

the sensor system (10) further comprises a confidence measure module arrangement used in combination with the at least one trained artificial neural network (16) to decide when to perform a decision or regression on the input data (x) of the at least one sensor (12) and when not, wherein the confidence measure module arrangement includes:

- a ΰf -module as implementation of a machine learning based method that is configured to learn a representation of the training dataset with trainable parameters f, wherein the representation of the training dataset is lower- dimensional than the training dataset, and

- an E-module as implementation of a measure to determine how far the input data (x) are from the training dataset using the information of

wherein

the ΰf -module is configured to provide for a

(x)-output of a module which learned the representation of the training dataset, said output being used by the E-module to determine how different the input data (x) are from what has been seen during training,

the E-module is configured to provide for an E(Ef (c),c)- output of a module which determines how far the input data (x) are from the training dataset using ΰy (x), said output being combined with M_g to decide whether the module is allowed to perform an action (classification or regression task), and the M_q -module is configured to provide for an M_q (x)-output of a classification or regression model using the input data (x),

wherein the confidence measure module arrangement is configured to utilize said E(Rf (c), c)- output and said M_q (x)- output in combination to decide to perform an action, upon which the output will be the classification or regression M_g(x) based on the input data (x), and

wherein the evaluation device (14) is configured to generate an output signal (20) representing a result of the classification or regression task.

2. The sensor system (10) according to claim 1 , wherein the at least one sensor (12) is formed as a radar sensor or as an optical camera.

3. The sensor system (10) of claim 1 or 2, wherein the machine learning based method implementation ΰ_f comprises a variational autoencoder.

4. The sensor system (10) according to claim 3, wherein the E-module is configured to calculate an /₂-error, l₂ between an

output of the autoencoder, denoted by Y, and the input data (x) as the measure to detect whether the input data (x) is close to the training dataset (module E(ΰ_f (c), c)).

5. The sensor system (10) according to one of claims 1 to 4, wherein in case that a value of the measure describing how far the input data (x) are from the training dataset Y is lower than a predefined threshold value, the M^-module is configured to perform a classification or regression based on the input data (x), otherwise, the M^-module is configured to generate a signal indicating that it is hindered to perform such action.

6. The sensor system (10) according to one of claims 1 to 5, wherein the ΰ_f- module and the M^-module are trained together/or in parallel and are configured to interact with each other.

7. The sensor system (10) according to claim6, wherein in case the

module and the M^-module are trained together, the M^-module is configured to use the input data (x) as well as the lower-dimensional representation Z of the input data (x) in order to perform a classification or regression task.

8. The sensor system (10) according to one of claims 1 to7, wherein the ΰ_f- module and the M^-module are trained independently from one another and are combined in the sensor system (10) after training.

9. The sensor system (10) according to claim 8, wherein the

module comprises a region of interest algorithm that is configured for proposing only predefined interesting regions in an image.

10. The sensor system (10) according to claim 9, wherein the region of interest algorithm is configured to be background-independent.

11. The sensor system (10) according to claim 10, wherein the M^-module is configured to perform a classification for each of the predefined interesting regions.

12. Method of operating a sensor system (10), the sensor system (10) including at least one sensor (12) and an evaluation device (14) that is configured for receiving input data (x) from the at least one sensor (12), and that comprises at least one trained artificial neural network (16) with trained parameters Q that is configured to perform a classification or regression task on data received from the at least one sensor (12), the method comprising at least the following steps:

- providing (30) input data (x) of the at least one sensor (12) to the evaluation device (14),

- applying (32) a machine learning based method implementation ΰ_f that has learned a representation of a training dataset with trainable parameters f, wherein the representation of the training dataset is lower-dimensional than the training dataset, to the input data (x) of the at least one sensor (12) for providing a W_f (x)-output,

- by using the information of W_f (x), applying (34) a measure implementation E for determining how far the input data (x) are from the training dataset and obtaining a measure output E(r_f (c), c),

- comparing (36) the obtained measure output E(r_f (c), c) with at least one predefined threshold value for the measure, - performing (38) a classification or regression task on the input data (x) received from the at least one sensor (12), if a result of the comparison fulfills a predefined condition, and

- generating (42) an output signal (20) representing a result of the classification or regression task, if a classification or regression task has been performed.

13. The method of claim 12, wherein the step of applying (34) a measure implementation E for determining how far the input data (x) are from the training dataset includes calculating an l₂-err or as a measure according to

14. The method of claim 12 or 13, further comprising the step (40) that, if the result of the comparison fails to fulfill the predefined condition, a signal is generated that indicates that the step (38) of performing a classification or regression on the input data (x) has been hindered.