DE102020214715A1 - NORMALIZED AUTOENCODER - Google Patents

Abstract

A computer-implemented method (10) for training a generative machine learning model, comprising: obtaining (12) multiple input samples in a data space, mapping (14) the input samples into a latent space as multiple samples in the latent space, reconstructing (16) multiple synthetic samples in the data space based on the latent space samples, and iteratively optimizing (18) the generative machine learning model based on an optimization function including a reconstruction loss and a test statistic to yield a plurality of model parameters.

Description

technical field

The present invention relates to a computer-implemented method for training a generative machine learning model, a computer-implemented method for generating synthetic data samples using a generative machine learning model, and associated devices. Furthermore, the present invention relates to an associated computer program element and an associated computer-readable medium.

background

The development and application of data-supported algorithms in technical systems is becoming increasingly important in the digitalization and especially in the automation of technical systems. In technical systems it can be advantageous to generate new data points, and in particular a large number of new data points. This allows, for example, various future scenarios to be simulated and statistically evaluated.

This subject area includes, for example, the generation of vehicle trajectory prediction data based on previously obtained vehicle trajectory data. An adversarial autoencoder (AAE) is a commonly used generative model. However, such approaches can be further improved.

Summary of the Invention

• - Erhalten mehrerer Eingangsproben in einem Datenraum;
• - Abbilden der Eingangsproben in einen latenten Raum als mehrere Proben in dem latenten Raum;
• - Rekonstruieren mehrerer synthetischer Proben in dem Datenraum auf der Grundlage einer oder mehrerer Proben des latenten Raums; und
• - iteratives Optimieren des generativen Maschinenlernmodells auf der Grundlage einer Optimierungsfunktion, die einen Rekonstruktionsverlust und eine Teststatistik umfasst, um mehrere Modellparameter zu liefern.

According to a first aspect, a computer-implemented method for training a generative machine learning model is provided. The computer-implemented method includes the following:

• - Obtaining multiple input samples in a data space;
• - mapping the input samples into a latent space as a plurality of samples in the latent space;
• - reconstructing a plurality of synthetic samples in the data space based on one or more samples of the latent space; and
• - iteratively optimizing the generative machine learning model based on an optimization function that includes a reconstruction loss and a test statistic to yield multiple model parameters.

According to a second aspect, an apparatus for training a generative machine learning model is provided. The facility includes an input interface configured to receive a plurality of input samples, a processor configured to receive the plurality of input samples, and an output interface configured to output a trained model and/or a plurality of synthetic data samples generated by the trained model.

The processor is configured, in a training phase, to map the plurality of input samples via a processor-executed encoder into a latent space as a plurality of samples of the latent space, to reconstruct a plurality of synthetic data samples via a processor-executed decoder based on one or more Samples the latent space, and to iteratively optimize, using the processor, the generative machine learning model based on an optimization function that includes a reconstruction loss and a test statistic.

• - Konfigurieren eines Decoder-Netzes gemäß den mehreren gemäß dem Verfahren des ersten Aspekts oder dessen Ausführungsformen erzeugten Modellparametern;
• - Erhalten mindestens einer Probe einer Generatorverteilung;
• - Decodieren der mindestens einen Probe der Generatorverteilung unter Verwendung des Decoder-Netzes, um somit weitere mehrere synthetische Datenproben in dem Datenraum zu erhalten; und
• - Ausgeben der weiteren mehreren synthetischen Datenproben.

According to a third aspect, a computer-implemented method for generating synthetic data samples using a generative machine learning model is provided. The procedure includes the following:

• - configuring a decoder network according to the plurality of model parameters generated according to the method of the first aspect or its embodiments;
• - obtaining at least one sample of a generator distribution;
• - decoding the at least one sample of the generator distribution using the decoder network, thereby obtaining further multiple synthetic data samples in the data space; and
• - Outputting the further multiple synthetic data samples.

According to a fourth aspect, there is provided an apparatus arranged to generate synthetic data samples using a generative machine learning model, where the generative machine learning model is optionally generated according to the method of the first aspect. The device comprises an input interface designed to receive a plurality of model parameters that may have been generated according to the method of the first aspect. The device includes a processor configured to obtain, based on a test statistic, one or more samples of the latent space configured according to the plurality of model parameters to obtain the decode one or more samples of the latent space to obtain another plurality of synthetic data samples in the data space, and output the other plurality of synthetic data samples.

According to a fifth aspect, there is provided a computer program element, comprising at least: (i) computer-executable instructions for training, using machine learning, a generative machine learning model according to the method of the first aspect and/or (ii) computer-executable instructions for generating synthetic data samples according to the Method of the third aspect and/or (iii) computer-executable instructions comprising model parameters for providing a generative machine learning model trained according to the method of the first aspect.

According to a sixth aspect there is provided a computer-readable medium comprising one or more of the computer program elements of the fifth aspect.

• - Erhalten mehrerer synthetischer Datenproben wie gemäß dem computerimplementierten Verfahren des ersten Aspekts oder dessen Ausführungsformen erzeugt;
• - Eingeben der mehreren synthetischen Datenproben in ein weiteres Maschinenlernmodell;
• - iteratives Optimieren des weiteren Maschinenlernmodells auf der Grundlage einer Optimierungsfunktion, die einen Rekonstruktionsverlust umfasst, um dadurch weitere mehrere Modellparameter zu aktualisieren; und
• - Ausgeben weiterer mehrerer Modellparameter des weiteren Maschinenlernmodells.

According to a seventh aspect, there is provided a computer-implemented method for training another machine learning model, comprising:

• - obtaining a plurality of synthetic data samples as generated according to the computer-implemented method of the first aspect or embodiments thereof;
• - inputting the plurality of synthetic data samples into another machine learning model;
• - iteratively optimizing the further machine learning model based on an optimization function including a reconstruction loss to thereby update further multiple model parameters; and
• - Outputting further several model parameters of the further machine learning model.

Accordingly, the present patent specification describes a new architecture, which is referred to here as a "normalized autoencoder" (NAE) and can be trained on input data, for example, in order to generate output data with a statistical distribution similar to the distribution D _i of the plurality of input data.

An adversarial autoencoder (AAE) typically includes an autoencoder with an additional Generative Adversarial Network (GAN). The GAN forces the distribution _Dl of the multiple data in the latent space of the adversarial _autoencoder to be similar to a chosen generator distribution Dg. Typically, to obtain new data, an AAE is trained without supervision on the available input data. A sample is then generated from the distribution D _g and passed through the AAE's trained decoder. This results in a new output data point with a similar appearance to the input data on which the AAE was trained.

According to one technique of the present specification, the architecture of a normalized autoencoder (NAE) results from replacing the GAN of an AAE with a test statistic S. The test statistic evaluates the distribution of the multiple samples in the latent space for similarity to a distribution D or distribution class to identify C. In the latter case, it also returns information about the distribution in C that is closest to _Dl . The test statistic requires no training, yet provides useful information about an expected property of _Dl .

GANs require a neural network, and thus a significant training complexity, which in practice means that a large number of vector and matrix multiplication and addition operations are performed one after the other. On the other hand, many test statistics for the distribution D can be calculated with a significantly lower running time, for example using a smaller number of vector multiplications. Accordingly, one effect of replacing the GAN of an AAE with a test statistic, as in the NAE presented here, is a reduction in computational complexity, and thus runtime and/or power consumption when generating new data.

In addition, the resulting model generated by the NAE disclosed in aspects herein has higher fidelity compared to that of an AAE. In the case of an AAE trained using a GAN, the distribution D is typically fixed. Alternatively, the test statistic S, as used in the NAE discussed in aspects herein, can be configured to only test on the plurality of data in the latent space that are in a given distribution class C. For example, the test on the multiple data in the latent space that is in a given distribution class could be independent of the mean and/or the covariance of the multiple data in the latent space. In this case, the mean and covariance of the multiple data in the latent space remained flexible, making the NAE easier to optimize.

In addition, GANs, such as those found in AAEs, are difficult to train and often exhibit erratic training behavior. Through the proposed use of the test statistic S instead of the GAN in the optimization term according to the aspects discussed here simplifies the training of the AAE architecture, since a test statistic S generally has no trainable parameters. Thus, the training process is reduced to the training process that should be performed for the autoencoder without introducing additional stability and training problems through the GAN of an AAE.

The approach of the present specification relates to a method for training a decoder neural network of an NAE on multiple input data samples having the distribution D _i , so as to provide a decoder capable of providing synthetic data samples that closely match the distribution D _i hold the multiple input data. In other words, multiple model parameters of at least one decoder network capable of generating the multiple synthetic data samples are provided.

The training of at least the NAE's decoder network is affected by a two-component loss function. The first component of the loss function is the reconstruction loss (L), obtained by comparing input data samples encoded into latent space with corresponding reconstructed output samples decoded out of latent space. The second component of the loss function is obtained by applying the test statistic S to multiple samples (Z) in the latent space. The test statistic evaluates how similar/close the distribution of its input samples (the latent space samples in this application) is to distribution D or a distribution in distribution class C.

In other words, the NAE's test statistic evaluates whether the samples at its input form a distribution that has a shape similar to the shape it is testing for. For example, the test statistic can evaluate the kurtosis of an input sample set. For example, the kurtosis of a univariate normal distribution is 3. In one example, the test statistic replaces the discriminator of a generative adversarial network.

The approach of the present patent also relates to a method for generating a plurality of synthetic data samples using the trained neural decoder network according to the plurality of model parameters of at least one decoder network calculated in the training phase. Typically, when generating synthetic samples, the decoder takes samples from the chosen generator distribution D _g , decodes the samples obtained from the generator distribution and outputs synthetic samples with a distribution similar to that of the multiple input data D _i used to train the NAE, out.

Another way to look at the technique discussed here is to consider the different types of distribution. An input data distribution D _i characterizes the distribution of multiple input samples X={x ₁ , x ₂ , x ₃ ,..., x _n }. A distribution D ₁ in the latent space characterizes the distribution of the latent space samples Z={z ₁ , z ₂ , z ₃ ,..., z _n } when trained by a trained encoder model. A decoding by a trained decoder model yields several synthetic samples X̂ = (x̂ ₁ , x̂ ₂ , x̂ ₃ , ..., x̂ _n }. Their distribution approximates the distribution D _i of the several input data X = {x ₁ , x ₂ , x ₃ , ..., x _n } A comparison of X with X̂ yields a loss function that can be used in the optimization (back propagation) of the (neural network of) the encoder and/or the decoder model.

The present technique also proposes an addition to the loss function. This is obtained by evaluating the distribution D _l of samples Z = (z ₁ , z ₂ , z ₃ , ..., z _n } of the latent space using a test statistic S. The test statistic can test, for example, to what extent D _l resembles a Gaussian distribution For example, if applying the test statistic S to D ₁ reveals that the distribution of the multiple samples Z={z ₁ , z ₂ , z ₃ ,..., z _n } of the latent space is not comparatively Gaussian -type, this is reflected in the loss function.The encoder and/or the decoder model is adjusted over a number of training iterations to match the distribution of the multiple samples Z = {z ₁ , z ₂ , z ₃ , ..., z _n } of the latent space lead to a more Gaussian-type shape (in one example).

In a data generation mode of the NAE, a decoder trained under the influence of the test statistic S is given random samples either from the distribution D (if S tested with respect to D) or from the distribution in the distribution class C that was closest to _Dl (if S with respect to the class C tested). Accordingly, the distribution of further multiple synthetic samples X̂ _s = {x̂ _s1 , x̂ _s2 , x̂ _s3 , ..., x̂ _sm } is similar to that of the multiple input data X = {x ₁ applied to the encoder and decoder models in the training phase , x2 , _x3 , ..., _{xn }} .

character list

• 1A FIG. 12 schematically illustrates a computer-implemented method for training a generative machine learning model according to a first aspect.
• 1B Figure 12 schematically illustrates a computer-implemented method for generating synthetic data samples using a generative machine learning model according to a third aspect.
• 2 shows the training of a generative machine learning model schematically.
• 3 Figure 12 schematically depicts a facility for generating synthetic data using a machine learning model.
• 4 shows a device according to the second aspect schematically.

Detailed description

An improvement to adversarial autocoders (AAEs) is now presented. In particular, a Normalized Autoencoder (NAE) is presented, which can be used at least in data generation scenarios where generative models are currently used. An important application of generative models is, for example, the generation of training data or test data from existing training or test data. This finds application, for example, in anomaly detection where there is little training data comprising training samples with an anomaly compared to training samples without an anomaly.

An anomaly can be, for example, a sensor or other hardware component of a machine that is malfunctioning. Another example where research data is not readily available in the required amounts is autonomous driving applications. Obtaining realistic data of unsafe autotrajectories is more difficult in practice than obtaining realistic data of safe autotrajectories. Accordingly, the present specification discusses at least one method for generating training data, a method for generating test data to verify that a trained model is safe to operate, and a generative model for generating the training or test data.

An autoencoder (AE) is a neural network architecture consisting of an encoder neural network and a decoder neural network designed to generate reconstruction loss. The encoder neural network is designed to map from a data space to a latent space. In one example, the latent space has a smaller dimension than the data space. The decoder neural network is designed to map from latent space back to data space. The reconstruction loss is calculated. Reconstruction loss is a measure of similarity between a given element of data space before and after reconstruction from latent space. The weights of the encoder neural network and the decoder neural network are trained by minimizing the loss between the initial data and the data after it has passed through the autoencoder. In one example, the function Avg(L(data, DEC(ENC(data)))) is minified.

A further development of the autoencoder is the adversarial autoencoder, which includes an additional generative adversarial network. The additional generative adversarial network introduces an additional loss term. The additional loss term is, in one example, proportional to the deviation of the distribution of the encoded input data (data in the latent space) from a predefined distribution.

In other words, the additional loss term of an AAE approximates the so-called Jensen-Shannon divergence between the two distributions. By minimizing the additional loss term, the AAE's encoder is forced to encode the input data in such a way that the data is distributed in the latent space approximately according to the predefined distribution.

The output of the training process (the autoencoder) can be provided as a generative model for the data distribution, for example by inputting random samples taken from a generator distribution (D _g ) into the decoder after training the AAE. To generate synthetic data from the generative model, a random sample of the latent distribution is taken. The random sample is decoded using the decoder trained in the training phase. The random sample has approximately the same distribution as the original input data.

This property of adversarial autoencoders results from the fact that the data in the latent space is distributed approximately according to the predefined distribution and sampling from the predefined distribution is relatively easy to set up. However, adversarial autoencoders can be further improved.

• - Erhalten 12 mehrerer Eingangsproben in einem Datenraum;
• - Abbilden 14 der mehreren Eingangsproben in einen latenten Raum als mehrere Proben in dem latenten Raum Z;
• - Rekonstruieren 16 mehrerer synthetischer Proben in dem Datenraum auf der Grundlage der mehreren Proben in dem latenten Raum Z; und
• - iteratives Optimieren 18 des generativen Maschinenlernmodells auf der Grundlage einer Optimierungsfunktion, die einen Rekonstruktionsverlust und eine Teststatistik umfasst, um mehrere Modellparameter zu liefern.

According to a first aspect, there is provided a computer-implemented method 10 for training a generative machine learning model, comprising:

• - obtaining 12 multiple input samples in a data space;
• - mapping 14 the plurality of input samples into a latent space as a plurality of samples in the latent space Z;
• - reconstructing 16 a plurality of synthetic samples in the data space based on the plurality of samples in the latent space Z; and
• - iteratively optimizing 18 the generative machine learning model based on an optimization function that includes a reconstruction loss and a test statistic to yield multiple model parameters.

1A shows a computer-implemented method for training a generative machine learning model according to the first aspect schematically.

The general approach presented here is to replace the generative adversarial network used in an AAE with a test statistic S. Therefore, according to a first aspect, a model is provided for generating test data by using a neural network, such as an autoencoder, to minimize the test statistic, rather than minimizing the Jensen-Shannon divergence as a GAN would do.

2 Figure 12 schematically illustrates an example of training a generative machine learning model according to an example device, in particular a Normalized Autoencoder, NAE. A NAE comprises an encoder network 13a, a decoder network 15a, a test statistics engine 28 and a reconstruction loss engine 30. The Encoder network 13a is arranged to map input data received from an input interface 12a in the data space into a latent space memory 14a.

In one example, latent space memory 14a stores latent space samples generated by encoder 13a.

In one example, the encoder network 13a and/or the decoder network 15a provides one, two, three or more hidden layers. In one example, latent space store 14a stores vectors having a lower dimensionality compared to the input data received from input interface 12a.

The decoder network 15a is configured to map samples generated and/or stored by the latent space store 14a back into the data space as reconstructed data provided to an output interface 18a of the generative machine learning model. In one example, encoder network 13a and/or decoder network 15a are configured as deep neural networks, although this is not a requirement.

The reconstruction loss engine 30 is configured to compare, in data space, an input data item provided by the input interface 12a with a corresponding output data item provided by the output interface 18a. The reconstruction loss engine 30 produces the reconstruction loss L, a measure of how similar these two elements of the data space are. For example, in a successful training operation, the reconstruction loss should be reduced (backpropagation) with each iteration and backpropagation performed, for example, as part of an iterative gradient descent optimization process.

The test statistics engine 28 receives its input from the latent space vector z maintained in the latent space store 14a. The test statistics engine 28 outputs a value S(z).

In one example, the test statistic S is evaluated with respect to all samples in latent space Z. In one example, the test statistic S is evaluated with respect to a subset of the samples in the latent space Z.

In one example, the test statistic engine 28 outputs a relatively small value when multiple input samples z from the latent space store 14a appear to have been taken from the latent distribution D (or from one of the distributions in C) according to the selected test statistic S. In one example, the test statistic engine 28 returns a relatively large value when multiple input samples z do not appear to be samples of the latent distribution D (or any distributions in C) according to the chosen test statistic S.

In one example, the output of the test statistic engine 28 returns a value close to zero when a plurality of input samples z appear to be a sample of the latent distribution D (or a sample from a distribution in C) according to the chosen test statistic S, and close to one if several input samples z do not appear to be samples of the latent distribution D (or any distributions in C) according to the chosen test statistic S.

In this example, the test statistic engine 28 is configured to generate a test statistic S as a measure of how much the distribution of the data in the latent space differs from a distribution D, or distribution class C. In other words, the test statistic engine 28 is designed to generate a test statistic 28 that provides a mapping from a set of points in the latent space to real numbers. In order to train the normalized autoencoder (NAE), in a generation mode, input samples X = {x ₁ , x ₂ , x ₃ , ..., x _n } are obtained from a data source to be emulated by the NAE. In one example, encoder 13a, decoder 15a, and/or the hidden layer are initialized to zero, a set of random numbers, or an arbitrary initialization vector. In one example, the initialization condition of the nets is specified as a hyperparameter.

For a first iteration, a reconstruction loss L and a test statistic S are calculated. The reconstruction loss is calculated based on the loss between the input samples X and the encoded and then decoded input samples X . The test statistic S is calculated based on a calculation of the encoded value of the input samples X and a calculation of a value of the test statistic based on the result.

The total loss is calculated in block 32 by combining the reconstruction loss L and the test statistic S.

On subsequent iterations, the NAE is trained by minimizing L(X, DEC(ENC(X))) + S(ENC(X)).

In one example, the reconstruction loss L at each iteration is a function of the multiple input samples X = {x ₁ , x ₂ , x ₃ , ..., x _n ) and the synthetic multiple samples X̂ = {x̂ _s1 , x̂ _s2 , x̂ _s3 , ..., x̂ _sn } (during each iteration). The test statistic S(z) is a function of at least one sample of the latent space Z={z ₁ , z ₂ , z ₃ ,..., z _n }.

The minimization is performed, for example, by adjusting the weights in the encoder and decoder networks. For example, the gradient descent method can be used. After a number of iterations, or until a convergence criterion is met, as appropriate, the decoder 15a weights form a decoder model that can be used as a generative model for generating further arbitrary sequences of synthetic output samples similar to the input samples.

In one example, S(z) takes on small values when the samples of latent space Z appear to be samples of the distribution against which the test statistic is testing, and takes on large values otherwise.

In one example, the test statistic S(z) is a metric that measures a similarity or dissimilarity between an empirical distribution of the plurality of input samples Z={z ₁ , z ₂ , z ₃ ,..., z _n } and a predefined distribution or a predefined distribution class defined.

In one example, the test statistic S(z) is minimized when a sample of the latent space Z at least approximates a sample of the predefined distribution D or a distribution in the predefined distribution class C for which the test statistic is optimized.

In one example, the output value of the test statistic S(z) is based on at least one sample of the latent space Z.

In one example, the test statistic S(z) tests one or more latent space samples as to whether they are extracted from a normal distribution, a multivariate normal distribution, a multivariate exponential distribution, or a discrete random variable distribution in the latent space.

In one example, the test statistic S(z) characterizes the extent to which samples of the latent space or class of predefined distributions approximate a given shape, but are independent of the mean or variance of the predefined distribution or class of predefined distributions.

As an example, the test statistic S(z) is the test statistic for multivariate normality W _n,β according to the article "A New Approach to the BHEP Tests for Multivariate Normality" by Henze and Wagner, published in Journal of Multivariate Analysis 62 , pp. 1-23 (1997), published by Academic Press.

In one example, the mapping of the multiple input samples into the latent space as multiple samples in the latent space is performed by an encoder 13a, and the reconstructing of multiple synthetic samples into a data space from the latent space samples is performed by a decoder 15a.

In one example, iteratively optimizing the generative machine learning model based on an optimization function comprising the reconstruction loss L and the test statistic S includes iteratively adjusting multiple weights in the encoder 13a and/or the decoder 15a based on a combination of the reconstruction loss L and the test statistic S.

In one example, the generative machine learning model is a Normalized Autoencoder.

In one example, the test statistic replaces an adversarial generative network of an adversarial autoencoder. Accordingly, the training process of the normalized autoencoder benefits from a higher degree of stability. In one example, the test statistic identifies a kurtosis of a statistical distribution.

In one example, the generative machine learning model is output to and/or stored in memory.

According to a second aspect, a device 34 for training a generative machine learning model is provided. The facility includes an input interface 36 configured to receive a plurality of input samples, a processor 38 configured to receive the plurality of input samples, and an output interface 42 configured to output a trained model and/or a plurality of synthetic data samples generated by the trained model.

The processor 38 is configured, in a training phase, to map the multiple input samples via an encoder 13a executed by the processor into a latent space as multiple samples in the latent space, to reconstruct multiple synthetic data samples via a decoder 15a executed by the processor 38 based on the latent space samples, and iteratively optimizing, using the processor, the generative machine learning model based on an optimization function that includes a reconstruction loss and a test statistic.

4 shows a device according to the second and/or the fourth aspect schematically.

In one example, device 34 is a personal computer, a server, a cloud-based computer, or an embedded computer. The memory 40 of the device 34 stores a computer program which, when executed by the processor 38, causes the processor 38 to carry out the functionalities described by the computer-implemented methods according to the first or the third aspect.

According to one example, the input interface 36 and/or the output interface is a USB interface, or an Ethernet interface, or a WLAN interface, or other suitable hardware that can enable the input and output of data samples from the device 34 .

In one example, the facility 34 further comprises a volatile and/or non-volatile storage system 40 configured to receive input observations as input data from the input interface 40, for example to store a test statistic S or a plurality of latent spaces.

• - Konfigurieren 20 eines generativen Maschinenlernmodells gemäß mehreren Modellparametern;
• - Erhalten 22, auf der Grundlage einer Teststatistik, einer Probe des latenten Raums aus gemäß den mehreren Modellparametern konfigurierten latenten Variablen;
• - Decodieren 24 der Probe des latenten Raums, um weitere mehrere synthetische Datenproben in dem Datenraum zu erhalten; und
• - Ausgeben 26 der weiteren mehreren synthetischen Datenproben.

According to a third aspect, there is provided a computer-implemented method 11 for generating synthetic data samples using a generative machine learning model, comprising:

• - configuring 20 a generative machine learning model according to a plurality of model parameters;
• - obtaining 22, based on a test statistic, a sample of the latent space from latent variables configured according to the plurality of model parameters;
• - decoding 24 the latent space sample to obtain further multiple synthetic data samples in the data space; and
• - Output 26 of the other multiple synthetic data samples.

3 Figure 12 schematically depicts a facility for generating synthetic data using a machine learning model.

Following the training process described above in connection with the first aspect, synthetic data samples can be obtained that have a distribution D that approaches or is similar to that of the data used to train the NAE. Sampling an example, a decoder network is configured according to the decoder model parameters calculated according to the first aspect or its embodiments. A generator distribution engine 29 is sampled by a sampler 31 to provide input samples to the trained decoder 15a'. The prime character indicates that the decoder 15a has been configured using model parameters calculated in a previous training iteration.

The decoder 15a' receives the samples from the generator distribution engine 29 and decodes them, yielding multiple synthetic data samples X̂ _s = {x̂ _s1 , x̂ _s2 , x̂ _s3 , ..., x̂ _sn }. The multiple synthetic data samples have a statistical distribution that approximates the distribution of the data X used to train the NAE.

In other words, the generator distribution engine 29 can be understood as replacing the latent space Z samples from the training phase at the input of the trained decoder 15a'.

By sampling using the generator distribution engine 29 to provide surrogate latent space samples and computing DEC(z) using the trained decoder network 15a', a data sample of synthetic output data X̂ _s = {x̂ _s1 , x̂ _s2 , x̂ _s3 , ..., x̂ _sn } that closely resemble the data used to train the NAE.

In one example, when S is testing against a distribution class C, the parameters of C's closest matching element D(n) are estimated before latent space Z is scanned. Then the best matching element D(n) is used to sample Z.

• - eine Eingabeschnittstelle 36, ausgelegt zum Empfangen mehrerer Eingangsproben;
• - einen Prozessor 38, ausgelegt zum Erhalten der mehreren Eingangsproben; und
• - eine Ausgabeschnittstelle 42, ausgelegt zum Ausgeben eines trainierten Modells und/oder mehrerer durch das trainierte Modell erzeugter synthetischer Datenproben; wobei der Prozessor ausgelegt ist, in einer Trainingsphase, zum Abbilden der mehreren Eingangsproben über einen durch den Prozessor ausgeführten Encoder in einen latenten Raum als eine oder mehrere Proben des latenten Raums Z, zum Rekonstruieren mehrerer synthetischer Datenproben über einen durch den Prozessor ausgeführten Decoder 15a auf der Grundlage der Proben des latenten Raums; und
• - zum iterativen Optimieren, unter Verwendung des Prozessors, des generativen Maschinenlernmodells auf der Grundlage einer Optimierungsfunktion, die einen Rekonstruktionsverlust und eine Teststatistik umfasst, um mehrere Modellparameter zu liefern.

According to a fourth aspect, a device 34 for training a generative machine learning model is provided. The facility includes the following:

• - an input interface 36 adapted to receive a plurality of input samples;
• - a processor 38 arranged to receive the plurality of input samples; and
• - an output interface 42 arranged to output a trained model and/or a plurality of synthetic data samples generated by the trained model; wherein the processor is adapted, in a training phase, to map the plurality of input samples via a processor-executed encoder into a latent space as one or more samples of the latent space Z, to reconstruct a plurality of synthetic data samples via a processor-executed decoder 15a the basis of the latent space samples; and
• - to iteratively optimize, using the processor, the generative machine learning model based on an optimization function that includes a reconstruction loss and a test statistic to yield multiple model parameters.

According to a fifth aspect, there is provided a computer program element, comprising at least: (i) computer-executable instructions for training, using machine learning, a generative machine learning model according to the method of the first aspect or examples thereof, and/or (ii) computer-executable instructions for generating synthetic Data samples according to the third aspect and/or (iii) computer-executable instructions comprising model parameters for providing a generative machine learning model trained according to the first aspect.

The computer program elements of the fifth aspect may, for example, comprise machine-readable instructions storable in a computer memory.

According to a sixth aspect there is provided a computer-readable medium comprising one or more of the computer program elements of the fifth aspect.

In one example, the plurality of input samples X comprise one or more instances of system anomalies, and the plurality of synthetic data samples Xs generated by the synthetic data generator using a machine learning model comprise a plurality of synthetic system anomalies. An anomaly is a pattern that deviates from an expected result. Accordingly, a generative machine learning model trained according to the first aspect is designed to generate a series of synthetic system anomalies. These synthetic system anomalies can in turn be used to train another machine learning model (e.g. an anomaly detector designed to detect system anomalies in an input stream of system data). For example, the input samples may be obtained from a sensor (e.g., a thermal monitor) of a machine (e.g., an engine), and a system anomaly may indicate a predetermined condition of the machine (e.g., a fault condition, such as an engine overheating condition based on a previous engine speed). The synthetic system anomalies can be used to train another machine learning model configured to detect a state of the machine based on an input stream of system data.

In other examples, the input samples X include one or more examples of data representing internal states of a technical system (e.g., a machine). The data representing internal states can be determined on the basis of sensor data and/or simulated for the respective technical system. The multiple synthetic data samples Xs generated by the device for generating synthetic data using a machine learning model can include multiple synthetic examples of data, which in these examples represent internal states of a technical system. Accordingly, a generative machine learning model trained according to the first aspect is designed to generate synthetic examples of data that represent internal states of a technical system. The synthetic examples of internal states The data representing the technical system can be used to train a further machine learning model (e.g. a machine learning model for monitoring and/or controlling the technical system).

Additionally or alternatively, the input samples X include one or more examples of data representing the environment of a technical system (e.g. a machine). The data representing the environment can be determined on the basis of sensor data and/or simulated for the respective technical system. The multiple synthetic data samples Xs generated by the synthetic data generating facility using a machine learning model may include multiple synthetic examples of data, which in these examples represent the environment of a technical system. Accordingly, a generative machine learning model trained according to the first aspect is designed to generate synthetic examples of sensor data that represent the environment of a technical system. The synthetic examples of data representing the environment of the technical system can be used to train a further machine learning model (e.g. a machine learning model for monitoring and/or controlling the technical system).

Additionally or alternatively, the input samples X comprise one or more instances of sensor data and the multiple synthetic data samples Xs generated by the synthetic data generator using a machine learning model comprise multiple synthetic instances of sensor data. The sensor data can be obtained, for example, for monitoring and/or controlling a technical system (e.g. a machine). Accordingly, a generative machine learning model trained according to the first aspect is designed to generate synthetic samples of sensor data. The synthetic examples of sensor data can be used to train another machine learning model (e.g. a machine learning model for monitoring and/or controlling the technical system).

In still other examples, the input samples include instances of image or video data (e.g., a vector or array of semantically segmented image data, e.g., pixel data). The multiple synthetic data samples Xs generated by the synthetic data generator using a machine learning model may include multiple synthetic samples of image or video data. The image or video data may be obtained from one or more sensors (e.g., a camera-based sensor, an ultrasonic sensor, a RADAR sensor, a LIDAR sensor, or a thermal sensor) and/or be synthetic image data. Accordingly, a generative machine learning model trained according to the first aspect is designed to generate synthetic samples of image or video data. The synthetic samples of image or video data can be used to train another machine learning model (e.g., a machine learning model for image or video data classification).

The input samples can be obtained from an industrial production machine or production line, for example. The input samples may include a vector of sensor inputs from all or a subset of the sensors included in the industrial manufacturing machine or assembly line. The sensors are designed to detect at least one of the following: the angular position of one or more servomotors, the power consumption of the one or more servomotors, the visual appearance of a portion of the industrial manufacturing machine or assembly line, feedback signals from multiple hydraulic or pneumatic actuators, and the like. In this case, a system anomaly may represent an object that is being processed incorrectly or is stuck in the industrial manufacturing machine or assembly line. In this case, the sensor signals will deviate in many respects from those expected in the normal manufacturing process. Additionally, a subset of sensor signals can provide advance notice of the anomaly. Therefore, the multiple synthetic data samples Xs include multiple synthetic sensor states of industrial manufacturing machines or assembly lines.

In one example, the multiple input samples X include one or more vehicle or robot trajectories in two- or three-dimensional space (e.g., Euclidean space), and the multiple synthetic data samples Xs generated by the synthetic data generator using a machine learning model include multiple synthetic vehicle trajectories . In the case of a vehicle, an anomaly may represent a vehicle veering into a wrong lane or following a trajectory that does not comply with traffic regulations. In the case of a robot, an anomaly can be a robot following a trajectory that results in a collision with an object in the robot's environment. Accordingly, another machine learning model configured with the model generated according to the first aspect is configured to generate a synthetic example of a plurality of synthetic vehicle trajectories.

In one example, the plurality of input samples X include one or more robotic arm trajectories in three-dimensional space (e.g., Euclidean space or polar space). In this case, an anomaly may be the robotic arm colliding with a stationary object within its range of motion, or the robotic arm attempting to move through a singularity (in the case of a multi-jointed robotic arm). The plurality of synthetic data samples Xs generated by the synthetic data generator using a machine learning model comprises a plurality of synthetic robotic arm trajectories.

In one example, the multiple input probes X include multiple probes that represent the state of a wireless communication network at a base station, a UE (User Equipment), or other element or network controller such as a RAN (Radio Access Node), an eNB ( Evolved Node B) or represent a gNB. For example, the multiple input samples X could include a figure of merit that summarizes the performance of the wireless network. For example, in a 5G or 4G system, the multiple input samples X might include one or more of the following: Adjacent Channel Power (ACP), EVM (Error Vector Magnitude) of received symbols, SNIR (Signal to Noise and Interference Ratio - signal-to-noise-plus-interference ratio), SNR (Signal to Noise Ratio), SS-RSRP (Synchronization Signal Reference Signal Received Power) and the like. An advantage of this is that a generative model that characterizes rare wireless network events can be provided to train another wireless network. Accordingly, another machine learning model configured with the model generated according to the first aspect is adapted to generate synthetic examples of multiple wireless network states or states of individual components of a wireless network.

It is not essential that the multiple input samples be scheduled based on acquisition time. For example, the input samples can be obtained based on the position of an actuator in three-dimensional space, the state of a phase plane, or the like.

In the cases discussed above, providing training data using the generative machine learning model according to the first aspect and its embodiments means that another machine learning model can be trained more efficiently and accurately. The further machine learning model (which is not limited to being an autoencoder) can be provided with input data for training that has realistic anomaly information that occurs far more frequently than anomalies observed in the normal operation of a given industrial system, autonomous vehicle, robotic arm or the like. Therefore, the further machine learning model converges faster and thus consumes less energy when training since candidate anomalies occur more frequently in training data provided according to the first aspect or its embodiments.

The examples provided in the drawings and described in the written description above are intended to provide an understanding of the principles of the disclosure. This is not intended to limit the scope of the disclosure. The present disclosure describes changes and modifications to the illustrated examples. Only the preferred examples have been presented, and any changes, modifications and further applications thereof within the scope of the disclosure are intended to be protected.

Claims (15)

A computer-implemented method (10) for training a generative machine learning model, comprising: - obtaining (12) a plurality of input samples (X) in a data space; - mapping (14) the input samples into a latent space as a plurality of samples in the latent space; - reconstructing (16) a plurality of synthetic samples in the data space based on one or more samples of the latent space (Z); and - iteratively optimizing (18) the generative machine learning model based on an optimization function comprising a reconstruction loss (L) and a test statistic (S) to yield multiple model parameters. Computer-implemented method (10) for training a generative machine learning model claim 1 , where the generative machine learning model is an autoencoder or a normalized autoencoder. Computer-implemented method (10) according to claim 1 or 2 , where the reconstruction loss on at least one iteration is a function of the multiple input samples and the synthetic multiple samples, and the test statistic is a function of at least one sample in the latent space (Z). Computer-implemented method (10) according to one of Claims 1 until 3 , wherein the test statistic is a metric that defines a similarity or dissimilarity of an empirical distribution of the plurality of input samples and a predefined distribution or a predefined distribution class. Computer-implemented method (10) according to claim 4 , further comprising: - minimizing the test statistic when a sample of the latent space at least approximates a sample of the predefined distribution or the predefined distribution class. Computer-implemented method (10) according to one of Claims 1 until 5 , where the value of the test statistic is based on at least one sample of latent space (Z). A computer-implemented method (10) as claimed in any preceding claim, further comprising: - testing, using the test statistic, the predefined distribution or the class of predefined distributions, to perform an evaluation for the existence of a normal distribution and/or a multivariate exponential distribution and/or a discrete random variable distribution. Computer-implemented method (10) according to claim 7 , further comprising: - iteratively optimizing the generative machine learning model based on an optimization function comprising the reconstruction loss and the test statistic by iteratively adjusting multiple weights in the encoder and/or the decoder based on a combination of the reconstruction loss and the test statistic. Computer-implemented method (10) according to one of the preceding claims, wherein the generative machine learning model is a normalized autoencoder and/or wherein the test statistic replaces a generative adversarial network of an adversarial autoencoder. A computer-implemented method (10) as claimed in any preceding claim, further comprising: - Obtaining, as the multiple input samples, one or more examples of system anomalies, and wherein the multiple synthetic data samples include multiple synthetic system anomalies, or Obtaining, as the multiple input samples, one or more vehicle trajectories and wherein the multiple synthetic data samples include multiple synthetic vehicle trajectories. A computer-implemented method (11) for generating synthetic data samples using a generative machine learning model, comprising: - configuring (20) a decoder network (15a') according to the plurality according to the method of any one of Claims 1 until 10 calculated model parameters; - obtaining (22) at least one sample of a generator distribution (D); - decoding (24) the at least one sample of the generator distribution (D) using the decoder network, thereby obtaining further multiple synthetic data samples in the data space; and - outputting (26) the further plurality of synthetic data samples. Apparatus (34) for training a generative machine learning model, comprising: - an input interface (36) arranged to receive a plurality of input samples; - a processor (38) arranged to receive the plurality of input samples; and - an output interface (42) arranged to output a trained model and/or a plurality of synthetic data samples generated by the trained model; wherein the processor (38) is adapted, in a training phase, to map the plurality of input samples via an encoder executed by the processor into a latent space as one or more samples of the latent space (Z); - for reconstructing, via a decoder executed by the processor, a plurality of synthetic data samples based on one or more samples of the latent space (Z); and - to optimize iteratively, using the processor, the generative machine learning model based on an optimization function that includes a reconstruction loss and a test statistic to yield multiple model parameters. One or more computer program elements comprising at least: (i) computer-executable instructions for training, using machine learning, a generative machine learning model according to the method of any one of Claims 1 until 10 and/or (ii) computer-executable instructions for generating synthetic data samples claim 11 and/or (iii) computer-executable instructions that provide the model parameters len one after one of the Claims 1 until 10 trained generative machine learning model. Computer-readable medium carrying one or more of the computer program elements Claim 13 includes. A computer-implemented method (11) for training another machine learning model, comprising: - obtaining a plurality of synthetic data samples as according to the computer-implemented method of claim 11 generated; - inputting the plurality of synthetic data samples into another machine learning model; - iteratively optimizing the further machine learning model based on an optimization function comprising a reconstruction loss (L) to thereby update further multiple model parameters; and - Outputting a further number of model parameters of the further machine learning model.

Images