DE102021200215A1 - Determining a confidence of an artificial neural network - Google Patents

Abstract

Method for determining a confidence of a classifying ANN with the following steps: - specification (S1) of input data with which the ANN is fed to determine a classification result for the input data; - specification or generation (S2) of disturbed input data; - determination (S3 ) the classification result for the specified input data using the ANN;- determining (S4) the classification result for the noisy input data using the ANN;- determining (S5) the confidence in the classification result of the specified input data by comparing the classification result of the specified input data with the classification results of the noisy Input data is compared.

Description

FIELD OF THE INVENTION

The present invention relates to a method for determining a confidence of a classifying artificial neural network.

SUMMARY OF THE INVENTION

The present invention is based on the object of specifying the reliability of a classification result of a classifying artificial neural network.

• - Ein Verfahren zum Ermitteln einer Konfidenz eines klassifizierenden KNN mit den folgenden Schritten: Vorgeben von Eingangsdaten, mit welchen das KNN zum Ermitteln eines Klassifikationsergebnisses für die Eingangsdaten gespeist wird; Vorgeben oder Generieren von gestörten Eingangsdaten; Ermitteln des Klassifikationsergebnis für die vorgegebenen Eingangsdaten mittels des KNN; Ermitteln des Klassifikationsergebnis für die gestörten Eingangsdaten mittels des KNN; Ermitteln der Konfidenz des Klassifikationsergebnis der vorgegebenen Eingangsdaten, indem das Klassifikationsergebnis der vorgegebenen Eingangsdaten mit den Klassifikationsergebnissen der gestörten Eingangsdaten verglichen wird.

Accordingly, it is provided:

• - A method for determining a confidence of a classifying ANN with the following steps: specification of input data, with which the ANN is fed to determine a classification result for the input data; specifying or generating noisy input data; Determining the classification result for the specified input data using the ANN; determining the classification result for the noisy input data using the ANN; Determining the confidence of the classification result of the specified input data by comparing the classification result of the specified input data with the classification results of the noisy input data.

An artificial neural network (ANN) is in particular a network of networked artificial neurons that is simulated in a computer program. The artificial neurons are typically arranged on different layers. Usually, the artificial neural network comprises an input layer and an output layer (output layer), whose neuron output is the only one of the artificial neural network that is visible. Layers lying between the input layer and the output layer are typically referred to as hidden layers. Typically, an architecture or topology of an artificial neural network is first initiated and then trained in a training phase for a specific task or for a plurality of tasks in a training phase.

An ANN defines a mapping from an (example two-dimensional) input data space to a result data space. The mapping depends significantly on the purpose of the artificial intelligence, for example it is conceivable to classify data points of the input data space with regard to their properties. In this case, the data points from the input data space are assigned a classification result, such as "tree" or "house" from the result data space, and one speaks of a classifying ANN.

The term "topology of an ANN" includes all aspects relating to the structure of an ANN. This includes, for example, the number of neurons in the ANN, the allocation of the neurons to the individual layers of the ANN, the number of layers in an ANN, the networking of the neurons and the weighting of the networking.

The training of the artificial neural network typically includes changing a weight of a connection between two artificial neurons of the artificial neural network. The weight contains information about the strength of the consideration of an input of a neuron. The training of the artificial neural network can also include developing new connections between artificial neurons, deleting existing connections between artificial neurons, adjusting threshold values of the artificial neurons and/or adding or deleting artificial neurons.

An input data space is a set of data that contains all conceivable or well-defined input data for an ANN. A result data space is a set that contains all conceivable result data of an ANN. This patent application is based on the assumption that an input data space can be partitioned into subsets or data points, with each element, ie a subset or a data point, of the partition being mapped to a different mapping result from the result data space. Boundaries of the subsets are also referred to as decision boundaries. Typical input data are, for example, n-dimensional vectors.

In this patent application, disturbed input data is input data that has been artificially linked to a disturbance.

Noise is a disturbance variable with a broad, non-specific frequency spectrum. It can therefore be interpreted as a superposition of many harmonic oscillations or waves with different amplitudes and frequencies or wavelengths. There are noise processes of different properties, e.g. white noise, pink noise or brown noise.

The amplitude of a one-dimensional disturbance is also referred to as its intensity. Is it a multidimensional Stö The disturbance intensity can be defined differently depending on the vector, eg as the magnitude of a disturbance vector or as the maximum of its components.

Confidence indicates a probability that a statement is correct. In common parlance, terms such as reliability or dependability are also used.

A primary use of an ANN is to apply it to a purpose for which the ANN is designed. This does not include maintenance, training or testing of the ANN. If the ANN is a driver assistance function, the primary use is assisted driving of a vehicle by an end user.

Input data specifically selected to determine data pairs from input data and an associated confidence may have been collected during a primary use of an ANN. Alternatively, such input data can also be artificially generated data.

A measure is a function that assigns numbers to subsets of a base set that can be interpreted as the sizes of those sets.

A sensor, also known as a detector, (measuring variable or measurement) pickup or (measuring) probe, is a technical component that detects certain physical, chemical properties or states, e.g. B. temperature, humidity, pressure, speed, brightness, acceleration, pH, ionic strength, electrochemical potential and / or the material composition of its environment can be qualitatively or quantitatively recorded as a measurand. These variables are recorded using physical or chemical effects and converted into an electrical signal that can be further processed as sensor data. Vehicle sensors are mounted on a vehicle to sense a vehicle environment. Sensor data collected from vehicle sensors are data points related to the vehicle environment.

Computer program products typically include a sequence of instructions that, when the program is loaded, cause the hardware to perform a specific method that leads to a specific result.

The basic idea of the invention is to assess the confidence of a trained classification system. Confidence approximates the distance from a data point in the input data space to the nearest decision boundaries. This allows an assessment of how close a data point is to another decision boundary. If it turns out that the distance between the data point in the input data space and another decision limit is small, the confidence in the classification result is low, since even a small variation in the input data would lead to a different classification result.

The assessment of the significance or reliability of a classification result of an artificial neural network is relevant for a large number of safety-critical applications. For example, it is relevant for a classifying artificial neural network in the training phase whether the confidence of an ANN has already reached an acceptable value or whether a human would generate better or worse classification results. Thus, for example, the progress during training can be assessed with regard to the reliability of the classification results. The present invention makes this possible by means of the method described below for determining a confidence of a classifying artificial neural network with automated methods based on an algorithm.

The method according to the invention is based on the assumption that the classification certainty, ie the confidence, of an artificial neural network is higher when the input data are in the center of a decision limit. At this point, the data is most reliably assigned to a specific classification result, and small changes in the data do not change the classification result. Accordingly, the measure of confidence should be such that at such points, the determined confidence of the artificial neural network is high. Accordingly, the confidence of a classification result of input data that is close to a decision boundary is low. A small change in the data can be enough to change the classification result of the ANN. Accordingly, the degree of confidence should be dimensioned in such a way that the determined confidence for input data is low in the vicinity of a decision limit.

In practice, however, the position of the decision limits is usually not known.

The present invention provides for determining the confidence of a classification result of a classifying artificial neural network by specifying input data for which the ANN determines a classification result. In the following, disturbed input data are generated or specified for the specified input data. In the following it is determined how close the specified input data is to a decision limit, in that the ANN determines classification results for the specified input data and for the disturbed input data. If the classification result of the specified input data differs from the disturbed input data, the result is that the disturbance was sufficient to transfer the specified input data to a different decision limit. This means that the distance between the given input data and a decision limit is smaller than the magnitude of the disturbance. Various methods, which can be weighted or unweighted, are available for comparing the classification results of the specified and disturbed input data.

Advantageous refinements and developments result from the further dependent claims and from the description with reference to the figures of the drawing.

According to a preferred development of the invention, a predetermined number of disturbed input data is generated by means of a noise process of at least one probability distribution.

Accordingly, it is conceivable to automatically generate disturbed input data for predetermined input data by means of noise. Provision can be made here for the disturbed input data to be generated by means of a noise process which is based on one or more probability distributions. All distributions come into consideration as probability distributions, for example the normal distribution, the uniform distribution, the Poisson distribution and the like.

Alternatively, it is also conceivable to specify deterministic disturbance vectors, by means of which the specified input data are linked. It can thus be ensured that the disturbed input data are comparable or are changed with regard to controllable properties.

According to a preferred development of the invention, one or more noise intensities are specified and for each noise intensity a predetermined number of disturbed input data is generated by means of a noise process of at least one probability distribution.

It is thus not only possible to assess whether specified input data is close to a decision limit, but also to assess the distance between the input data and the edge of the decision limit.

For all variants of the method described above and below, it is conceivable to implement the comparison of the classification results using a (normalized) sum of zeros or ones, with each classification result of noisy input data being compared once with the classification result of the specified input data, i. H. if the classification result of the noisy input data is identical to the classification result of the specified input data, a 1 is added and if the classification result of the noisy input data differs from the classification result of the specified input data, a 0 is added. This sum is determined for all classification results, ie all disturbed input data for all noise intensities.

In other words, the confidence of a classification result can then be formulated as follows, where i, for example, indexes different fault directions and j indexes different fault intensities: $$\mathrm{\xi }=\frac{1}{\text{N}\cdot \text{M}}{\displaystyle {\sum }_{\text{i},j}\left(y_0==y_{\text{i}+\text{j}\cdot \text{N}}\right)};\text{where i}\in \mathrm{1,}...,\text{N};\text{j}\in \mathrm{0,}...,\text{M}-1;$$

According to a preferred development of the invention, the specified input data are based on sensor data, with the sensor data being recorded for the purpose of a primary use of the artificial neural network.

This embodiment can also be referred to as a live variant in technical jargon. In this variant, the confidence is determined at the same time as the classification results of input data. Accordingly, the disturbed input data is also generated in real time and the classifying artificial neural network is fed with this disturbed input data for the purpose of its classification. Accordingly, the confidence of a classification result is determined while the classification result itself is determined.

This variant of the method can be easily integrated into existing systems of classifying ANN, since the existing classifying ANN does not have to be changed for this, but only has to be linked to a routine for determining the confidence.

This variant is particularly suitable for applications with large available computing capacities, as this requires additional computing effort.

Alternatively, it is also conceivable that the specified input data are used specifically to determine data pairs from input data and a con were selected. Accordingly, the procedure is carried out before the classifying ANN is used in its primary use.

It can be provided that a further ANN determines data pairs from input data and an associated confidence, while the classifying artificial neural network is fed with the specified input data based on sensor data, the sensor data being recorded for the purpose of a primary use of the ANN, where the further ANN was trained with data pairs from input data and an associated confidence.

Accordingly, it is provided that a further artificial neural network is trained prior to the primary use of the classifying ANN to assess the confidence of classification results during the primary use of the classifying ANN. The further ANN is trained before the classifying ANN is used in its primary use. For this purpose, data pairs from input data and an associated confidence are specified. During the primary use of the classifying ANN, the further ANN is then trained to determine the confidence of the classification results. This variant can also be referred to as a training variant in technical jargon.

For this variant, it is advantageous to change the classifying ANN by adding another output value, namely the confidence. It is also conceivable to train the further ANN together with the classifying ANN in order to keep the training effort low.

Alternatively, it is also conceivable to estimate the confidence of classification results of input data based on sensor data, the sensor data being recorded for the purpose of primary use of the ANN, by interpolating multiple data pairs from input data and an associated confidence. For this it is basically possible to use an identical data set as one would use for the training of another ANN, i.e. data pairs from input data and an associated confidence, whereby the data pairs are used specifically to provide a data set for the interpolation of confidences during the primary use of the KNN have been provided. This variant can also be referred to as the offline variant in technical jargon.

Accordingly, the offline variant is based on comparable or identical data to the training variant, with the offline variant being a simplification compared to the training variant. For this it is necessary to store data pairs from input data and an associated confidence. If, during the primary use of the classifying ANN, a confidence of a classification result of given input data for which no associated confidence was previously determined is to be determined, then this confidence can be determined by interpolating previously known confidences from the stored dataset.

It can be seen that the classifying ANN does not have to be changed for this and that it is therefore possible to expand any existing classifying ANN with the offline variant. Compared to the live variant or the training variant, the additional runtime requirement is lower, since only simple or fewer arithmetic operations are required.

According to a preferred development of the invention, the confidence is also determined based on a measure of a difference between the classification results of the noisy and no-noise input data if the classification results of the noisy and no-noise input data differ. Accordingly, in addition to the distance of an input data point to the next decision limit, the distance of a data point to any decision limit can also be estimated. For this it is advantageous to assess the distance to other decision regions using different noise intensities. In this way, disturbed input data can also reach decision regions that are further away.

This is advantageous when objects from different classification results are expected to behave similarly or identically. For example, it is known that classifying ANNs, which are used for driver assistance applications, classify camera data of a vehicle with regard to different types of road users contained in the camera data. Accordingly, such applications classify detected objects as trucks, cars, pedestrians, bicyclists, mopeds, and the like. It can be uncritical if, instead of a cyclist, an electric scooter rider, colloquially also referred to as an e-scooter rider, is detected, since these two objects usually move at a similar speed or with a similar movement behavior. This can be taken into account by weighting weaker the classification results of noisy input data that differ from the classification result of the specified input data, if it is to be expected that the objects of the different classification results behave similarly or identically.

In this case, it can optionally also be provided that the similarity or dissimilarity of classification results is assessed in view of the current problem to be solved. For example, it can make sense to predict the behavior of a car on a federal highway or on a highway based on its performance, among other things, if it is provided that a car is to be classified on the basis of an assumed performance. On the other hand, it is not to be expected that the behavior of a car in a traffic-calmed area differs due to its performance.

According to a preferred development of the invention, the classifying ANN has a known topology. Accordingly, the confidence of a known or specially developed classifying ANN is assessed. Accordingly, the confidence determined can be made accessible to a user, as a result of which he can assess the reliability of the system himself.

Alternatively, it is also conceivable that the classifying ANN has an unknown topology and confidences in classification results of the classifying ANN are to be determined or verified using the method described.

This makes it possible, for example, to assess the confidence of subordinate systems purchased for manufacturers of superordinate systems even if the topology of a subordinate system that has been purchased is not known. It is thus possible to determine the confidence of a classifying ANN even if its internal functioning is unknown and cannot be analyzed.

Confidence that was determined using the method described is also referred to as environmental confidence. Confidence determined in a different way or in an unknown way is also referred to as manufacturer confidence.

Alternatively and/or additionally, it is thus also possible to analyze environmental confidences of classifying ANNs with unknown topologies or to compare them with manufacturer confidences.

If this comparison shows a high correlation between the manufacturer's confidence and the environment's confidence, it could be concluded that the manufacturer's confidence is a reliable measure of the quality of classification results. Accordingly, however, it could also be concluded that the manufacturer confidence is not trustworthy or in need of improvement if there is a low correlation between the manufacturer confidence and the environment confidence.

A computer program product according to a method of an embodiment of the invention performs the steps of a method as described above when the computer program product runs on a computer, in particular an in-vehicle computer. If the program in question is used on a computer, the computer program product produces an effect, namely the assessment of the reliability of classification results of artificial neural networks.

character list

• 1 ein schematisches Blockdiagramm einer Ausführungsform der Erfindung;
• 2 eine Prinzipskizze zur Erläuterung einer Ausführungsform der Erfindung;
• 3 eine Prinzipskizze zur Erläuterung einer Ausführungsform der Erfindung.

The present invention is explained in more detail below with reference to the exemplary embodiments given in the schematic figures of the drawings. They show:

• 1 a schematic block diagram of an embodiment of the invention;
• 2 a schematic diagram to explain an embodiment of the invention;
• 3 a schematic diagram to explain an embodiment of the invention.

The accompanying drawings are provided to provide a further understanding of embodiments of the invention. They illustrate embodiments and, together with the description, serve to explain principles and concepts of the invention. Other embodiments and many of the foregoing advantages will become apparent by reference to the drawings. The elements of the drawings are not necessarily shown to scale with respect to one another.

In the figures of the drawings, elements, features and components that are identical, have the same function and have the same effect--unless stated otherwise--are each provided with the same reference symbols.

DESCRIPTION OF EXEMPLARY EMBODIMENTS

1 shows a schematic block diagram of a method for determining a confidence of a classifying artificial neural network according to an embodiment of the invention with the steps S1 to S5. In the step S1 input data with which the ANN is fed to determine a classification result for the input data is specified.

In step S2, disturbed input data are specified or generated.

In step S3, a classification result for the specified input data is determined using the ANN.

In step S4, classification results for the noisy input data are determined using the ANN.

In step S5, the confidence of the classification result of the input data is determined by comparing the classification result of the input data with the classification results of the noisy input data.

2 shows a schematic diagram to illustrate an embodiment of the invention. In 2 an input data space 19, which is broken down into a partition 13, is shown. The partition 13 comprises three subsets 15, 16 and 17. The subsets 15, 16 and 17 are delimited from one another by a subset boundary 11 which forms a decision boundary. A region 12 close to the edge 11 of the subset is shown in an illustrated manner within the subset 15 . Although the subset edge 12 is only shown for the subset 15, it goes without saying that the subsets 16 and 17 also have a comparable subset edge within these subsets. The subset 15 contains a plurality of data points 1 which are located within the subset 15 at a distance from the edge 12 of the subset. Furthermore, the subset 15 includes data points 2 that are located within the subset 15 and within the subset boundary 12 .

Subset 16 includes multiple data points 3. Subset 17 includes multiple data points 4.

A classifying ANN, which processes data from the input data space 19, would assign an object “house” to data points 1 and 2 of subset 15, an object “tree” to data points 3 of subset 16 and an object “tree” to data points 4 of subset 17. Assign vehicle". If there are data points in a close range 12 at the edge of the subset 11, their classification result, e.g. house, has a low confidence, since a small amount of interference is sufficient to bring the classifying ANN to a different classification result, e.g. tree.

3 schematically illustrates another embodiment of the invention. 3 shows a detailed view of the partition 13 according to FIG 2 . In 3 a data point 2, which is within the subset 15 and within the vicinity 12, is shown. Data point 2 has been disturbed by disturbance 5, which moves data point 2 from subset 15 to subset 16. Accordingly, disturbance 5 transforms data point 2 into data point 3 within subset 16.

Clearly explained using the above example, this means that the specified data point 2 z. B. was classified as a house and the noisy data point 3 was classified as a tree, for example. An embodiment of the invention makes it possible to assign a low confidence to the classification result “house” of data point 2 if the disturbance 5, which transferred data point 2 to another decision region, ie subset of the partition, was low.

Reference List

S1-S5

process steps

1-4

data points

5

Disturbance

11

subset edge

13

partition

15-17

subsets

19

input data space

Claims (11)

Method for determining a confidence of a classifying ANN with the following steps: - Specification (S1) of input data with which the ANN is fed to determine a classification result for the input data; - Specification or generation (S2) of disturbed input data; - Determining (S3) the classification result for the specified input data using the ANN; - Determining (S4) the classification result for the noisy input data using the ANN; - Determining (S5) the confidence of the classification result of the specified input data by comparing the classification result of the specified input data with the classification results of the disturbed input data. procedure after claim 1 , A predetermined number of disturbed input data being generated by means of a noise process of at least one probability distribution. Method according to one of the preceding claims, wherein one or more noise intensities are specified and a predetermined number of disturbed input data is generated for each noise intensity by means of a noise process of at least one probability distribution. Method according to any one of the preceding claims, wherein the predetermined input data is based on sensor data, the sensor data being collected for the purpose of a primary use of the ANN. A method according to any of the foregoing Claims 1 - 3 , where the given input data was specifically selected to determine data pairs from input data and an associated confidence. procedure after claim 5 , wherein a further ANN is provided for determining data pairs from input data and an associated confidence, while the classifying ANN is fed with predetermined input data based on sensor data, the sensor data being recorded for the purpose of a primary use of the ANN, the further ANN with data pairs from input data and an associated confidence claim 5 was trained. procedure after claim 5 , wherein the confidence of classification results of input data based on sensor data, wherein the sensor data were recorded for the purpose of a primary use of the ANN, by means of an interpolation of multiple data pairs from input data and an associated confidence claim 5 is appreciated. A method according to any one of the preceding claims, wherein the confidence is further determined based on a measure of a difference between classification results of the noisy and non-noiseed input data when the classification results of the noisy and non-noisy input data differ. A method according to any one of the preceding claims, wherein the classifying ANN has a known topology. A method according to any of the foregoing Claims 1 - 8th , wherein the classifying ANN has an unknown topology and confidences of classification results of the classifying ANN are determined and/or verified using the method according to one of the preceding claims. Computer program product with program code means to carry out the method according to one of the preceding claims.

Images