DE102019211672A1 - Training method for an artificial neural network - Google Patents

Abstract

Training method for an artificial neural network, which is set up to determine output data on the basis of input data, with the following steps: providing an output training data set which is provided with information, and an artificial neural network which was trained using the output training data set; Providing a target data set which differs from the initial training data set with regard to a characteristic; Training the artificial neural network using the target data set.

Description

FIELD OF THE INVENTION

The present invention relates to a training method for an artificial neural network and a driver assistance system with an artificial neural network.

SUMMARY OF THE INVENTION

The present invention is based on the object of reducing the effort involved in training an artificial neural network.

According to the invention, this object is achieved by a training method for an artificial neural network with the features of claim 1.

• - ein Trainingsverfahren für ein künstliches neuronales Netzwerk, welches eingerichtet ist, aufgrund von Eingangsdaten Ausgangsdaten zu ermitteln, mit den folgenden Schritten: Bereitstellen eines Ausgangstrainingsdatensatzes, welcher mit Informationen versehen ist, sowie eines künstlichen neuronalen Netzwerks, welches mittels des Ausgangstrainingsdatensatzes trainiert wurde; Bereitstellen eines Zieldatensatzes, welcher sich hinsichtlich einer Charakteristik von dem Ausgangstrainingsdatensatz unterscheidet; Trainieren des künstlichen neuronalen Netzwerks mittels des Zieldatensatzes.

Accordingly, it is provided:

• a training method for an artificial neural network which is set up to determine output data on the basis of input data, with the following steps: providing an output training data set which is provided with information, and an artificial neural network which was trained using the output training data set; Providing a target data set which differs from the initial training data set with regard to a characteristic; Training the artificial neural network using the target data set.

An artificial neural network (ANN) is in particular a network of networked artificial neurons, which is simulated in a computer program. The artificial neurons are typically arranged on different layers. The artificial neural network usually comprises an input layer and an output layer, the neuron output of which 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 special task or for several tasks in a training phase.

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

A training system is a computing unit on which an ANN is trained.

In this application, training data are data pairs of input data that are to be processed by the ANN, as well as target result data that are to be determined by the ANN. During the training, the ANN is adapted on the basis of a comparison of target result data with the actual result data determined by the ANN, which results in a training effect.

One possibility to train artificial neural networks is error feedback. A network makes mistakes in the early training phase. To correct the error, the triggers of the deviations (errors) between the generated assignment (actual output data) and solution assignment (target output data) are tracked and repeatedly controlling factors (weights) in the layers of the network are changed so that the assignment errors become smaller.

In this application, an initial training data record is a labeled training data record, i.e. a training data record which is provided with information with regard to semantic features that are contained in the data.

Labeled data or data that are provided with information are data records that have been provided with one or more pieces of information. Typically, a set of unlabeled data is used, supplemented with meaningful information. For example, labels can indicate whether a photo includes a horse or a cow, what words were spoken in an audio recording, what type of action is performed in a video, what is the subject of a news article and what the general mood is like, etc.

Labels are often generated by humans, e.g. by humans judging a certain piece of unlabeled data (e.g. "Does this photo contain a horse or a cow?"). As a result, labeled data is considerably more expensive than unlabeled raw data.

In this patent application, a target data set is a scenario on which an ANN is to be trained on the basis of an initial training data set. Accordingly, the differs Target data set in a characteristic, that is to say a characteristic property that is common to all data of the target data set, of the initial training data set.

GAN (Generative Adversarial Networks) are two artificial neural networks that perform a zero-sum game. One of them creates candidates (the generator), the second neural network evaluates the candidates (the discriminator). The generator typically maps a vector of latent variables to the desired result space. The goal of the generator is to learn to produce results according to a certain distribution. The discriminator, on the other hand, is trained to distinguish the results of the generator from the data from the real, given distribution. The generator's objective function is then to produce results that the discriminator cannot distinguish. As a result, the generated distribution should gradually adjust to the real distribution.

The distance indicates a difference between two elements of a data structure. A simple example of a distance measure is the distance between two points. However, a distance measure can be specified for every data structure that has at least two comparable elements.

A distance relation arranges elements, here data records, of a data structure with regard to the amount and their characteristics.

An intermediate data set is a data set that has a characteristic, the amount of which is between the amount of the characteristic of a starting training data set and a target data set. If the initial training data set comprises daylight images and the target data set night images, an intermediate data set can contain twilight images. The characteristic in this example is the brightness or light intensity.

An image file is a file that stores a digital image. The content of the file was either calculated digitally or digitized by analog-digital conversion and can therefore neither be recognized as an image nor read as text by the human viewer. A device that converts the content back into analog data is required for visualization. A distinction is made between raster graphics, which are composed of picture elements (pixels), and vector graphics, which are described with graphic primitives.

Intensity values can be specified for any number of measured values. Gray values are also intensity values. Brightness values are e.g. luminous intensity values and indicate the luminous flux related to the solid angle.

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 invention and further embodiments thereof are described below on the basis of a time-varying characteristic. It goes without saying that the invention can also be applied to characteristics that are not dependent on time, but on another parameter or change as a function thereof.

The invention is based on the assumption that input data to be evaluated for an artificial neural network change continuously. Input data to be evaluated are often related to the environment of a person, an object or a device. Accordingly, the present invention can be used, for example, for continuously changing surroundings or environmental data.

An example of a continuously changing environmental feature is brightness. The brightness increases continuously from dawn to noon, until it continuously decreases with dusk and dusk.

Another example of continuously changing input data for an artificial neural network are dialects or spoken dialects, which have to be analyzed by speech recognition software, since training data with different spoken dialects can be put into a certain order so that the voice recordings to be analyzed can be analyzed from a training data set to continuously change to a next training data set. For example, it is conceivable that an initial training data set for training an artificial neural network is spoken in High German and a target data set is spoken in Tyrolean dialect. It is conceivable to provide one or more intermediate data records that represent a step-by-step transition from High German to Tyrolean.

The invention is based on the aspect of the continuous transition between training data. The basic idea of the invention is to provide an artificial neural network which is trained on an initial training data set, as well as the initial training data set. In relation to the above examples, the initial training data set can be, for example, as Image data set with daylight recordings or as an audio data record with voice recordings in High German.

The invention is based on the assumption that the artificial neural network is sufficiently trained on the output training data set so that it determines output data suitable for corresponding input data.

The aim of the invention is to train the artificial neural network, which is trained on the initial training data set, on a target data set, the target data set of the initial training data set being spaced apart with regard to a characteristic. Accordingly, the initial training data set differs from the target data set, at least with regard to one characteristic.

The invention provides that the initial training data set is provided with information, i. That is, that the initial training data set is labeled. If the initial training data set is, for example, image data of a vehicle environment, this means that environmental features in the image data, such as a road, a motor vehicle, a cyclist, a lane marking, a crash barrier and the like, are noted in the image data. Because of this identification, the artificial neural network is trained for a wide range of situations that may result from possible features. Providing a training data set with information (labels) is usually time-consuming.

The invention makes it possible to use a lower level of detail of information in the target data record and / or intermediate data record. The invention even makes it possible not to provide a target data record and / or intermediate data record with information at all if the distance between the initial training data record and the neighboring data record, that is to say the target data record or an intermediate data record, is small enough.

In order to further train the artificial neural network which has already been trained on the target data set, it is sufficient to train the neural network on unlabeled input data of the target data set. The training of an artificial neural network can thus be simplified.

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

It is also expedient here if, with regard to the characteristic, there is a distance measure which defines a distance between the initial training data record and the target data record. A particularly clear distance measure is, for example, the distance between data of the MNIST training data set as the initial training data set and data of the MNIST training data set, which are rotated by 180 °, as the target data set. In this case, the distance between the initial training data set and the target data set is 180 °. It goes without saying, however, that distance measures between an initial training data set and a target data set are also conceivable that are less clear. For example, a distance measure can also be constructed which indicates the distance between two voice recordings.

The MNIST (Modified National Institute of Standards and Technology) database is a large database of records of handwritten digits that is commonly used to train various vision systems. An example from the MNIST database is in 1 to see.

It goes without saying that in the light of the invention it is also possible to define a distance measure which only defines a comparison relation larger or smaller between training data sets. For example, it is conceivable to train an artificial neural network for a driver assistance system with image data relating to the vehicle environment from 2010. It can be assumed that ten years later, i.e. in 2020, vehicle models will have changed in such a way that specific training on new vehicle models may be necessary. In this case, it is not absolutely necessary to quantify the distance between the data sets from 2010 and 2020. However, it is possible to determine that a further possible intermediate data set from 2015 has a smaller gap to the target data set than the initial training data set from 2010.

Consequently, it is also conceivable to define a distance measure or a distance relation for time-variable training data sets on the basis of the acquisition time.

According to a preferred development of the invention, at least one intermediate data set is generated or provided which, with regard to the characteristics, is arranged between the initial training data set and the target data set and the artificial neural network is trained by means of the intermediate training data set.

In relation to the above example, the intermediate data record can be used, for example, as a training data record with image data relating to the vehicle environment from the year 2015 if the initial training data record has vehicle-environment-related image data from 2010 and the target data record has vehicle-environment-related image data from 2020. In the above example, the properties of vehicle models form the characteristic.

With reference to the MNIST example, it can also be provided that the step size between the initial training data set and the next intermediate data set or between two adjacent intermediate data sets or between an intermediate data set and the target data set is 1 °. If the target data set is rotated by 180 ° compared to the initial training data set, 179 intermediate data sets are required with a step size of 1 °.

Accordingly, it can be provided that, with regard to the characteristic, a step size, which is the distance or a distance relation between the initial training data set and a first intermediate data set, a first intermediate data set and a second intermediate data set and / or an n-th intermediate data set and the target data set, where n is a is a positive integer, specifies, is determined and the artificial neural network is trained by means of the intermediate data sets.

Accordingly, it can be provided that the artificial neural network is trained step-by-step using intermediate data records until the intermediate data records have sufficiently approximated the target data record.

It goes without saying that it can also be provided that a suitable step size is determined, for example by means of an algorithm.

It can also be provided that the artificial neural network is trained, starting with the intermediate data set, which has the smallest distance to the initial training data set in terms of characteristics, and is then trained with the next adjacent intermediate data set or the target data set.

Using the example of the MNIST data set, this means, clearly explained, that the artificial neural network is initially trained with a first intermediate data set, which is rotated by 1 °, for example, with respect to the initial training data set. Next, the artificial neural network is trained with 178 further intermediate data sets, each of which is rotated by a further degree, until the artificial neural network has been trained with an intermediate data set that has been rotated by 179 ° compared to the initial training data set. As the last training of the artificial neural network, the training takes place by means of the target data set, which is rotated by 180 ° with respect to the initial training data set.

It can also be provided that the artificial neural network is retrained between this sequence using the initial training data set and / or using intermediate data sets that have already been trained. This ensures that the artificial neural network retains the capabilities with regard to the initial training data set or with regard to training data sets that have already been trained. If the artificial neural network forgets capabilities, this is also referred to in technical jargon as "catastrophic forgetting". This effect is reduced by retraining with data sets that have already been trained.

According to a preferred development of the invention, the target data record and / or the intermediate data record is provided with information in a lower degree of detail than the initial training data record. In particular, it can be provided that the target data record and / or the at least one intermediate data record are not provided with information. Data records that are provided with information are also referred to in technical jargon as labeled data records. The training of an artificial neural network with labeled data sets is also referred to in technical jargon as supervised machine learning. The training of an artificial neural network using unlabeled data sets is also known in technical jargon as unsupervised machine learning.

The labeling of training data sets is typically associated with a great deal of effort, so that monitored machine learning is significantly more complex and cost-intensive compared to unsupervised machine learning.

Due to the step-by-step training from a labeled initial training data set to a non-labeled or less labeled target data set, the positive effects of the monitored training of the initial training data set can be transferred to the subsequent, non-monitored training sessions. The training of the target or intermediate data record can thus be significantly improved, with the effort in terms of labeling being kept low.

As already described above on the basis of examples, the initial training data set, the target data set and the intermediate data set can comprise image data.

According to a preferred development of the invention, the characteristic is a time-variable parameter. It is provided that the characteristic of the initial training data set changes during a period of time to the characteristic of the intermediate data set or the target data set.

In this regard, it can be provided, for example, that the characteristic is a brightness value of image data.

Accordingly, it can be provided, for example, that an artificial neural network is trained by means of an initial training data record that includes image data that were recorded during the day. In the following, the target data record can for example include image data that were recorded during the night. Accordingly, intermediate data sets would gradually approach from daylight to nightlight. Accordingly, it can be provided, for example, that a possible intermediate data record includes image data that were recorded during twilight.

According to a preferred development of the invention, the artificial neural network that is trained using the initial training data set is stored as the first artificial neural network and a copy of the first artificial neural network is trained using the method as described above and as the second artificial neural network saved. In order to decide whether the first or the second artificial neural network is suitable for determining output data for specific input data, a classifier is used which is trained to determine the suitable artificial neural network.

Alternatively, it is also possible to train an artificial neural network in such a way that it determines suitable output data for input data in accordance with the output training data set, the target data set and optionally the intermediate data set.

According to a preferred development of the invention, the at least one intermediate data record and / or the target data record is generated by means of a GAN. Authentic training data can thus be generated by an algorithm. This is easier than collecting real training data.

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 brings about an effect, namely the simplification of training an artificial neural network.

Figure list

• 1 mehrere Trainingsdatensätze zum Trainieren eines künstlichen neuronalen Netzwerks mittels einer Ausführungsform der Erfindung;
• 2 mehrere Performanzdiagramme zu verschiedenen Ausführungsformen der Erfindung;
• 3 ein schematisches Blockdiagramm gemäß 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 several training data sets for training an artificial neural network by means of an embodiment of the invention;
• 2 several performance diagrams for various embodiments of the invention;
• 3 a schematic block diagram according to an embodiment of the invention.

The accompanying drawings are intended to provide a further understanding of the embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain principles and concepts of the invention. Other embodiments and many of the advantages mentioned emerge with a view to the drawings. The elements of the drawings are not necessarily shown to scale with one another.

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

DESCRIPTION OF EXEMPLARY EMBODIMENTS

1 shows training records 101 , 103 , 105 and 107 for training an artificial neural network according to an embodiment of the invention. The training data set 101 forms a labeled output training record containing handwritten numbers from the MNIST record. The record 107 forms a target data set and contains numbers of the MNIST data set which are opposite to the initial training data set 101 rotated by 90 °. The records 103 and 105 form intermediate data records, with the data record 103 is rotated by 30 ° with respect to the initial training data set and the training data set 105 is rotated by 60 ° compared to the initial training data set.

The aim of the procedure is to create an artificial neural network which is based on the recognition of the numbers of the initial training data set 101 is trained to recognize rotated numbers learn from the target data set and from the intermediate data sets. For this purpose, the artificial neural network is trained step by step, starting from the intermediate data records, until the rotation of the target data record has been reached.

Several coefficients can be set for training, which affect the performance of the method. The influence of the various coefficients is in 2 explained.

2 shows three diagrams 201 , 203 , 205 for the performance of a method which is trained to recognize numbers from the MNIST data set. The numbers from the MNIST data set are recognized by a trained artificial neural network.

The artificial neural network, its performance in the diagram 201 shown was trained 50 times on the MNIST output training dataset. In addition, the artificial neural network, its performance in the diagram 201 is shown trained twice on each intermediate record. In addition, the artificial neural network was applied once to the initial training data set in each intermediate step 101 retrained. In the diagram 201 the step size between training data sets is 1 °.

The curves 21st of the diagrams 201 , 203 , 205 each show the performance of the artificial neural network in the output training data set that is not rotated. The rotation to which the ANN was trained is plotted on the X-axis. The curves 22nd each show the performance of the artificial neural network for the intermediate data set, which is rotated by 10 ° with respect to the initial training data set. The curves 23 each show the performance of the artificial neural network for the intermediate data set, which is rotated by 20 ° with respect to the initial training data set. The curves 24 each show the performance of the artificial neural network for the intermediate data set, which is rotated by 30 ° with respect to the initial training data set. The curves 25th each show the performance of the artificial neural network for the intermediate data set, which is rotated by 40 ° with respect to the initial training data set. The curves 26th each show the performance of the artificial neural network for the intermediate data set, which is rotated by 50 ° with respect to the initial training data set. The curves 27 each show the performance of the artificial neural network for the intermediate data set, which is rotated by 60 ° with respect to the initial training data set. The curves 28 each show the performance of the artificial neural network for the intermediate data set, which is rotated by 70 ° with respect to the initial training data set. The curves 29 each show the performance of the artificial neural network for the intermediate data set, which is rotated by 80 ° with respect to the initial training data set. The curves 30th each show the performance of the artificial neural network for the intermediate data set, which is rotated by 90 ° with respect to the initial training data set.

All diagrams 201 , 203 and 205 show that the performance of an artificial neural network improves until the network has been trained to rotate the data set on which the performance is plotted. After that, the performance deteriorates again. For example, the performance of the curves increases 25th in each configuration until the artificial neural network has been trained to 40 °. After that, the performance decreases again in all configurations.

Hence goes from the diagram 201 shows that the performance on the initial training data set deteriorates somewhat as a result of the training of further intermediate data sets. In return, a significant improvement is achieved for all intermediate data sets and also for the target data set, but it can also be seen that the performance on the more strongly rotated intermediate data sets or on the target data set does not achieve approximately the same quality as that of the initial training data set. This effect can be seen in the diagram 201 may be attenuated by the training.

The artificial neural networks, their performance in the diagram 203 shown were trained 50 times on the initial MNIST training data set. In addition, the artificial neural networks, their performance in the diagram 203 is shown trained six times on each intermediate data set. Furthermore, the artificial neural networks were applied three times to the initial training data set in each intermediate step 101 retrained. In the diagram 203 the step size between training data sets is 1 °.

In contrast to the diagram 201 is in the diagram 203 , the performance on the initial training data set or on the early intermediate training data sets with increasing rotation of the intermediate data sets on which the ANN is trained, significantly worse. In return, however, the performance achieved on the intermediate data records is up to approx. 70 ° compared to the diagram 201 clearly improved.

The artificial neural networks, their performance in the diagram 205 shown were trained 50 times on the initial MNIST training data set. In addition, the artificial neural networks were trained once for each intermediate data set. During the training on the intermediate data sets, the artificial neural networks were not retrained on the initial training data set. The step size is 0.1 °.

Accordingly, the performance of the artificial neural networks between the individual training data sets is symmetrical. That is to say, the artificial neural network which is trained on the initial training data set starts with a very good performance for the initial training data set or with a poor performance for the target data set. At the end of the training, the performance is mirrored, i. H. the performance on the target data set is very good, corresponding to the performance on the initial training data set before the artificial neural network was trained on the intermediate data sets. In contrast, the performance on the initial training data set after the artificial neural network has been trained on the intermediate data sets and the target data set corresponds to the performance on the target data set before the artificial neural network has been trained on the target data set and the intermediate data sets.

3 FIG. 10 shows a block diagram according to an embodiment of a method for training an artificial neural network according to an embodiment of the invention. In the step S1 an initial training data set, which is provided with information, and an artificial neural network, which was trained using the initial training data set, are provided. In the step S2 a target data set, which differs from the initial training data set in terms of a characteristic, is provided. In the step S3 the artificial neural network is trained using the target data set.

List of reference symbols

21st

Performance curve

22nd

Performance curve

23

Performance curve

24

Performance curve

25th

Performance curve

26th

Performance curve

27

Performance curve

28

Performance curve

29

Performance curve

30th

Performance curve

101

Training data set.

103

Training data set.

105

Training data set.

107

Training data set

201

diagram

203

diagram

205

diagram

S1-S3

Procedural steps

Claims (15)

Training method for an artificial neural network, which is set up to determine output data based on input data, with the following steps: - providing (S1) an initial training data set which is provided with information, as well as an artificial neural network which was trained using the initial training data set; - Providing (S2) a target data set which differs from the initial training data set in terms of a characteristic; - Training (S3) the artificial neural network by means of the target data set. Training method for an artificial neural network according to Claim 1 , wherein with regard to the characteristic there is a distance measure which defines a distance or a distance relation between the initial training data set and the target data set. Training method for an artificial neural network according to one of the preceding claims, wherein at least one intermediate data set is generated or provided, which is arranged in terms of the characteristic between the initial training data set and the target data set and the artificial neural network is trained using the intermediate data set. Training method for an artificial neural network according to Claim 3 , wherein, with regard to the characteristic, a step size which is the distance between the initial training data record and a first intermediate data record, a first intermediate data record and a second intermediate data record and / or an n-th intermediate data record and the target data record, where n is a positive integer, is predetermined the artificial neural network is trained by means of the intermediate data sets. Training method for an artificial neural network according to Claim 4 , whereby a suitable step size between different data sets is determined. Training method for an artificial neural network according to one of the Claims 4 or 5 , wherein the artificial neural network is trained beginning with the intermediate data set, which has the smallest distance to the initial training data set in terms of characteristics, and is then trained with the next adjacent intermediate data set or target data set. The training method for an artificial neural network according to one of the preceding claims, wherein the artificial neural network according to the order according to Claim 6 is trained, the artificial neural network being retrained using the initial training data set and / or using an already trained intermediate data set. Training method for an artificial neural network according to one of the preceding claims, wherein the target data set and / or the intermediate data set is provided with information in a lower level of detail than the initial training data set, wherein in particular the target data set and / or the intermediate data set are not provided with information. The training method for an artificial neural network according to one of the preceding claims, wherein the initial training data set, the target data set and the intermediate data set comprise image data. Training method for an artificial neural network according to one of the preceding claims, wherein the characteristic is a time-varying parameter and the characteristic of the initial training data set changes over a period of time to the characteristic of the intermediate data set or the target data set. Training method for an artificial neural network according to Claim 10 , wherein the characteristic is a brightness value of image data and / o an intensity value. Training method for an artificial neural network according to one of the preceding claims, wherein the artificial neural network which is trained using the initial training data set is stored as the first artificial neural network and a copy of the first artificial neural network is trained using the method according to one of the preceding claims and is stored as the second artificial neural network and a classifier is trained which is set up to decide whether the first artificial neural network or the second artificial neural network is suitable for determining output data for specific input data. Training method for an artificial neural network, with at least one intermediate data set and / or the target data set being provided by means of a GAN. Driver assistance system with an artificial neural network which was trained with a training method according to one of the preceding claims. 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 1 - 13 be executed when the computer program product is running on the computer.

Images