DE102019205085A1 - Self-monitoring of a function based on artificial intelligence - Google Patents

Abstract

Method for monitoring first input data of an artificial intelligence function, which provides a core functionality and was trained with a training data set, by means of a monitoring function, which has a reconstruction function, which was trained with the training data set, the artificial intelligence function and the monitoring function respectively receive the same first input data, the reconstruction function mapping the first input data to condensed input data, which have a smaller dimension than the first input data, and reconstructing second input data from the condensed input data, and the monitoring function determining a reconstruction error between the first input data and the second input data.

Description

The present invention relates to a method for self-monitoring of a function based on artificial intelligence.

To reduce the risk of road traffic, assistance systems are known which record information about the surroundings by means of cameras and other sensors and recognize other vehicles therein. For example, the DE 102017210266 A1 such an assistance system.

Such assistance systems are typically based on an artificial intelligence function that has been trained with a training data set. During use, the problem can arise from this that an assistance system is confronted with a driving situation for which it has not been trained.

Against this background, the invention is based on the object of providing self-monitoring of an artificial intelligence function that recognizes application situations for which it has not been trained.

According to the invention, this object is achieved by a method for monitoring input data of an artificial intelligence function with the features of claim 1.

• - ein Verfahren zum Überwachen von ersten Eingangsdaten einer künstlichen Intelligenz-Funktion, welche eine Kernfunktionalität bereitstellt und mit einem Trainingsdatensatz trainiert wurde, mittels einer Überwachungsfunktion, welche eine Rekonstruktionsfunktion, welche mit dem Trainingsdatensatz trainiert wurde, aufweist, wobei die künstliche Intelligenz-Funktion und die Überwachungsfunktion jeweils gleiche erste Eingangsdaten erhalten, wobei die Rekonstruktionsfunktion die ersten Eingangsdaten auf kondensierte Eingangsdaten, welche eine geringere Dimension als die ersten Eingangsdaten aufweisen, abbildet und aus den kondensierten Eingangsdaten zweite Eingangsdaten rekonstruiert und wobei die Überwachungsfunktion einen Rekonstruktionsfehler zwischen den ersten Eingangsdaten und den zweiten Eingangsdaten ermittelt.

Accordingly, it is provided:

• - A method for monitoring first input data of an artificial intelligence function, which provides a core functionality and was trained with a training data set, by means of a monitoring function, which has a reconstruction function, which was trained with the training data set, the artificial intelligence function and the Monitoring function each receive the same first input data, the reconstruction function mapping the first input data to condensed input data, which have a smaller dimension than the first input data, and reconstructing second input data from the condensed input data, and the monitoring function a reconstruction error between the first input data and the second input data determined.

Vehicles within the meaning of this patent application are land, air and / or water vehicles.

An artificial intelligence function is based on artificial intelligence and is, for example, an artificial neural network. An artificial neural network is an intelligent algorithm. Intelligent algorithm means that the algorithm includes means of artificial intelligence, in English as artificial intelligence, abbreviated to AI.

The artificial neural network is implemented on an electronic circuit and is programmed based on the model of the neural network of the human brain. Functional units of an artificial neural network are artificial neurons whose output is generally evaluated as the value of an activation function using a weighted sum of the inputs plus a systematic error, the so-called bias. Artificial neural networks, similar to the human brain, are trained by testing several predetermined inputs, also called target training data, with various weighting factors and / or activation functions. Target training data are positive training data with which the artificial neural network learns true information. Target training data are labeled with the meaning of the information, that is, marked so that the artificial neural network can semantically capture information. Training with (un) marked training data is called (un) supervised learning.

The training of an artificial neural network with the help of target training data is called machine learning. A subset of machine learning is deep learning, so-called deep learning, in which a number of hierarchical layers of neurons, so-called hidden layers, are used to carry out the process of machine learning. An artificial neural network with several hidden layers is a deep neural network.

In the training phase, connections between neurons are assessed using weighting factors. Forward feeding, referred to as forward propagation, means that information is fed into the input layer of the artificial neural network, passes through the following layers and is output in the output layer. Reverse feeding, referred to in English as backward propagation, means that information is input into the output layer and output in the input layer. By successively feeding back the error from the output layer into the respective previous layer up to the input layer, the errors of the respective layers are preserved. The errors are a function of the weighting factors. By minimizing the error in the training phase, the weighting factors are changed. As a result, when the power is fed in again, the desired output is approached. The reverse feed is detailed in Michael A. Nielsen, Neural Networks and Deep Learning, Determination Press, 2015.

A trained artificial neural network is characterized by a purposeful reaction to new information.

Examples of artificial neural networks are convolutional or recurrent neural networks. Convolutional artificial neural networks are characterized by a two- or three-dimensional arrangement of neurons and divided weighting factors and are used in particular for image recognition in which the input is a gray image or an image in a three-dimensional color space and thus a two- or three-dimensional matrix. A recurrent, also feedback, neural network is a network that is characterized by connections between neurons in one layer and neurons in the same or a previous layer. Artificial recurrent neural networks are generally used for problems that require the processing of sequences. Examples are handwriting recognition, speech recognition, and machine translation.

Input data are, for example, sensor data that an artificial intelligence function receives by means of an interface from sensors.

This patent application distinguishes between a core function, for example the control of a vehicle, environment detection or trajectory planning, and a monitoring function by means of which the core function is monitored. Both the core function and the monitoring function are based on artificial intelligence.

A reconstruction function is trained to recognize structures in input data. The reconstruction function is set up to condense input data (first input data) based on their structure, i.e. to reduce the dimension of the input data, and to reconstruct input data from the condensed data (second input data) that are as similar as possible to the original input data (first input data). An example of a reconstruction function is an auto-encoder. An autoencoder is an artificial neural network that is used to learn efficient coding. The aim of an autoencoder is to learn a compressed representation (encoding) for a set of data and thus also to extract essential features. This means that it can be used to reduce dimensions.

Since the original input data (first input data) and the reconstructed input data (second input data) are typically not identical, a reconstruction error occurs when the data is reconstructed.

Clustered data is data that is grouped based on certain properties. For example, it can be provided that one or more parameters in a data cluster run through a specific interval.

An empirical (p) quantile, also simply called a quantile for short, is a term from statistics and a key figure for a sample. In simplified terms, an empirical p-quantile divides the sample in such a way that a portion of the sample of p is smaller than the empirical p-quantile and a portion of 1-p of the sample is greater than the empirical p-quantile.

In this patent application, additional safety precautions are understood to mean, for example, the reduction of a speed or a warning to a motor vehicle driver.

An interface is a device between at least two functional units at which an exchange of logical quantities, for example data, or physical quantities, for example electrical signals, takes place, either only unidirectionally or bidirectionally. The exchange can be analog or digital. The exchange can also take place in a wired or wireless manner.

A central data store is, for example, a cloud, i.e. a storage space that is provided as a service via the Internet.

Computer program products generally comprise a sequence of instructions which, when the program is loaded, cause the hardware to carry out a specific method that leads to a specific result. The computer program preferably comprises an architecture of an artificial neural network.

The core of the invention is to assess the similarity of current input data to original training data by means of a reconstruction function. The reconstruction function is trained to reconstruct second training data that is as similar as possible from first training data. The reconstruction function is accordingly suitable for reconstructing similar data from other data that are similar to the training data. Conversely, this means that if the reconstructed (second) input data are dissimilar to the first input data, the first input data are also dissimilar to the first training data.

If the first input data are dissimilar to the first training data, it can be concluded that there is a use case for which the artificial intelligence function was not designed.

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

According to a preferred development of the invention, it is detected that the artificial intelligence function is unsuitable for processing the first input data when the reconstruction error exceeds a predetermined threshold value. This is advantageous for safety-critical functionalities, for example, since it can thus be ensured that the artificial intelligence function does not leave the application area for which it was designed or trained unnoticed.

It can be useful to terminate the artificial intelligence function and / or to take additional safety precautions if the reconstruction error exceeds the predetermined threshold value. This prevents the artificial intelligence function from processing input data for which it was not trained. In security-critical applications, the processing of input data with an artificial intelligence function for which the function has not been trained is a security risk, since the artificial intelligence function is then no longer familiar with the application.

In addition, there may also be other reasons, for example economic reasons, to limit the range of applications of an artificial intelligence function.

According to a preferred development of the invention, the condensed input data are designed as clustered input data. This means that the clustered input data can also be compared with clustered training data. For example, the comparison can reveal that the reconstruction error corresponds to the distance between the first input data and the closest training data cluster.

According to a preferred development of the invention, a reconstruction error is determined for the training data set with first training data in that the reconstruction function maps the first training data onto condensed training data and reconstructs second training data from the condensed training data. It is provided that a reconstruction error is determined between the first training data and the second training data, the reconstruction error of the input data being assessed on the basis of the reconstruction error of the training data. According to a preferred development of the invention, the reconstruction function is designed as an auto-encoder. It goes without saying that other functions that rely on artificial intelligence are also suitable for reconstructing input data. Typically, such reconstruction functions are trained using unsupervised training methods (unsupervised learning).

According to a preferred development of the invention, the artificial intelligence function and in particular the reconstruction function are trained using the first training data when the reconstruction error in the input data exceeds the predetermined threshold value. In this case, the method detected that the artificial intelligence function was being applied to input data for which it was not originally designed. In this case, it can be advantageous to add these input data to a training data set so that the artificial intelligence function can be applied to this input data in the future.

According to a preferred development of the invention, the threshold value of the reconstruction error in the input data is determined based on a quantile of the distribution of the reconstruction errors in the training data. Accordingly, it can be provided that a distribution of the reconstruction errors in the training data is determined and a permissible reconstruction error in the input data is determined on the basis of a quantile of the distribution of the reconstruction errors in the training data.

According to a preferred development of the invention, the threshold value of the reconstruction error in the input data is established on the basis of a multiple of the largest reconstruction error in the training data. Accordingly, it can be provided, for example, that a largest permissible reconstruction error in the input data is 1.5 times the largest reconstruction error in the training data.

It goes without saying that an evaluation unit for monitoring input data, which is set up to carry out the method as described above, has an interface that can be connected to at least one sensor in order to receive the input data, a core function unit which is set up to carry out the artificial intelligence function, and with a monitoring unit which is set up to carry out the monitoring function, is advantageous.

In addition, there are central systems for generating training data, which can be connected to a large number of evaluation units as described above, and a central data memory and an interface between the Have a plurality of evaluation units and the central data memory, input data being transmitted from one evaluation unit of the plurality of evaluation units to the central system and stored in the central data memory if the evaluation unit has a reconstruction error in the input data while the method is being carried out as described above, which exceeds the predetermined threshold value is determined and the central system is set up to generate training data from the input data stored in the central data memory, advantageously.

In this way, input data for which an artificial intelligence function was not designed can be made available to a large number of evaluation units as training data. Thus, the range of applications of an artificial intelligence function can still be used during the period of use, i.e. i.e. after the appropriate software has been installed on a computer.

It goes without saying that vehicles with an evaluation unit as described above are advantageous.

The computer program product according to an embodiment of the invention carries out the steps of a method according to the preceding description when the computer program product runs on a computer, in particular a computer internal to the vehicle. When the program in question is used on a computer, the computer program product produces an effect, namely the monitoring of an artificial intelligence function.

• 1 ein schematisches Blockdiagramm einer Ausführungsform der Erfindung;
• 2 ein schematisches Blockdiagramm einer Ausführungsform der Erfindung;
• 3 ein schematisches Blockdiagramm einer Ausführungsform der Erfindung.

The present invention is explained in more detail below with reference to the exemplary embodiments specified in the schematic figures of the drawings. It shows:

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

1 shows a block diagram of a method for monitoring input data of an artificial intelligence function 14th . The block diagram shows a training phase 3 in which the artificial intelligence function 14th and the reconstruction function 16 with a training data set 12 be trained. The training phase takes place accordingly 3 takes place before the actual application of the method for monitoring input data.

The block diagram also shows an application phase 5 the process flow during the application. Both the artificial intelligence function 14th as well as the monitoring function 15th receive first input data 18th . The first input data 18th can for example from a sensor 34 are recorded. The artificial intelligence function 14th processes the first input data 18th to perform a core function such as controlling a vehicle. The monitoring function 15th monitors the input data 18th as to whether the artificial intelligence function 14th also for processing the first input data 18th is suitable, so whether the artificial intelligence function 14th on the first input data 18th was trained. The monitoring function assesses this 15th by means of the reconstruction function 16 the structural similarity between the training data set 12 and the first input data 18th . The reconstruction function forms 16 the first input data 18th on condense input data 20th from. This is the dimensionality of the input data 18th reduced. During the training phase 3 became the reconstruction function 16 trained on it, training data from the training data set 12 to condense with little loss and then to reconstruct. About the structural similarity of the first input data 18th to assess the training data reconstructs the reconstruction function 16 the condensed input data 20th to second input data 22nd .

The invention is based on the assumption that condensed input data 20th with a minor error on second input data 22nd can be reconstructed when the first input data 18th have a structure similar to that of the training data. The monitoring function determines accordingly 15th a reconstruction error 24 between the first input data 18th and the second input data 22nd . If the reconstruction error falls below 24 a specified threshold value, the reconstruction of the second input data 18th found to be good, ie that the first input data 18th are sufficiently similar to the training data. Consequently, it can be assumed that the artificial intelligence function 14th for processing the first input data 18th is suitable when the reconstruction error 24 falls below a predetermined threshold value.

2 shows an evaluation unit 30th according to one embodiment of the invention. The evaluation unit 30th includes an artificial intelligence function 14th as well as a monitoring function 15th . In order to receive input data, the evaluation unit comprises 30th an interface 32 that with a sensor 34 is connectable. The sensor 34 is in 2 trained as a camera.

3 shows a central system 42 with a central data store 40 , which is designed as a cloud and with a large number of evaluation units 30th is connectable.

List of reference symbols

3

Training phase

5

Application phase

12

Training data set

14th

artificial intelligence function

15th

Monitoring function

16

Reconstruction function

18th

first input data

20th

condensed input data

22nd

second input data

24

Reconstruction error

30th

Evaluation unit

32

interface

34

sensor

40

central data storage

42

Central system

QUOTES INCLUDED IN THE DESCRIPTION

This list of the 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.

Patent literature cited

• DE 102017210266 A1 [0002]

Claims (13)

Method for monitoring first input data (18) of an artificial intelligence function (14), which provides a core functionality and has been trained with a training data set (12), by means of a monitoring function (15) which has a reconstruction function (16) which was trained with the training data set (12), wherein the artificial intelligence function (14) and the monitoring function (15) each receive the same first input data (18), wherein the reconstruction function (16) maps the first input data (18) onto condensed input data (20) which have a smaller dimension than the first input data (18) and reconstructs second input data (22) from the condensed input data (20) and wherein the monitoring function (15) determines a reconstruction error (24) between the first input data (18) and the second input data (22). Method for monitoring first input data (18) according to Claim 1 wherein it is detected that the artificial intelligence function (14) is unsuitable for processing the first input data (18) if the reconstruction error (24) exceeds a predetermined threshold value. Method for monitoring first input data (18) according to one of the preceding claims, wherein the artificial intelligence function (14) is ended and / or additional safety precautions are taken if the reconstruction error (24) exceeds the predetermined threshold value. Method for monitoring first input data (18) according to one of the preceding claims, wherein the condensed input data (20) are in the form of clustered input data. Method for monitoring first input data (18) according to one of the preceding claims, wherein a reconstruction error (24) is determined for the training data set (12) with first training data by the reconstruction function (16) mapping the first training data onto condensed training data and from the condensed training data Training data reconstructs second training data and a reconstruction error (24) is determined between the first training data and the second training data, the reconstruction error (24) of the input data being assessed on the basis of the reconstruction error (24) of the training data. Method for monitoring first input data (18) according to one of the preceding claims, wherein the reconstruction function (16) is designed as an autoencoder. Method for monitoring first input data (18) according to one of the preceding claims, if on Claim 2 referred back, the artificial intelligence function (14) and in particular the reconstruction function (16) being trained using the first input data (18) if the reconstruction error (24) of the input data exceeds the predetermined threshold value. Method for monitoring first input data (18) according to one of the preceding claims, wherein the threshold value of the reconstruction error (24) of the input data is determined on the basis of a quantile of the distribution of the reconstruction errors (24) of the training data. Method for monitoring first input data (18) according to one of the preceding claims, wherein the threshold value of the reconstruction error (24) of the input data is defined on the basis of a multiple of the largest reconstruction error (24) of the training data. Computer program product which is designed to be loaded into a memory of a computer and which comprises software code sections with which the steps of the method according to one of the preceding claims are carried out when the computer program product runs on the computer. Evaluation unit (30) for monitoring first input data (18), which is set up to carry out the method according to one of the preceding claims an interface (32) which can be connected to at least one sensor (34) in order to receive the first input data (18), a core functional unit which is set up to carry out the artificial intelligence function (14) and with a monitoring unit which is set up to carry out the monitoring function (15). Central system (42) for generating training data, which is based on a plurality of evaluation units (30) Claim 11 can be connected, and has a central data memory (40) and an interface between the plurality of evaluation units (30) and the central data memory (40), first input data (18) from an evaluation unit (30) of the plurality of evaluation units (30) the central system (42) are transmitted and stored in the central data memory (40) if the evaluation unit during the Implementation of the method according to one of the Claims 1 - 9 a reconstruction error (24) of the first input data (18) which exceeds the predetermined threshold value is determined, and wherein the central system (42) is set up to generate training data from the input data stored in the central data memory (40). Vehicle comprising an evaluation unit according to Claim 11 .

Images