DE202021105192U1 - Apparatus for machine learning from a small number of training data examples - Google Patents

Abstract

Device for teaching a machine learning system (60),
wherein an input of the machine learning system (60) is suitable for receiving at least one pair of variables and _{processing them into an output variable (y pred} ) of the machine learning system (60),
wherein the device comprises a machine-readable storage medium on which instructions are stored which, when executed by a computer, caused the computer to carry out a method comprising the following steps:
Providing (S11) a first probability distribution (p (f)) which describes a probability distribution of a plurality of heuristics (f), the heuristics (f) outputting estimates of a parameter for a sensor variable;
Repeating the steps multiple times:
Randomly selecting (S12) one of the heuristics (f) as a function of the first probability distribution (p (f));
Generation (S13) of training data (D) as a function of the selected heuristic (f), a parameter being assigned to a plurality of sensor variables by means of the heuristic (f); and
Teaching (S14) the machine learning system (60) on the basis of the training data (D), the sensor variables and associated parameters being fed as pairs to the machine learning system (60).

Description

The invention relates to a device for teaching a machine learning system and for operating the learned machine learning system so that the machine learning system can output, for example, a classification from unknown classes from a small number of training data examples.

State of the art

Shi Xiahan, et al. "Relational generalized few-shot learning." ArXiv preprint arXiv: 1907.09557 (2019), available online: https://arxiv.org/abs/1907.09557, reveal a method which models and thus relationships between seen and new classes in the common label space enables generalized learning based on a small number of examples, also known as 'generalized few-shot learning'.

An example from the publication is Rasmussen, Carl Edward. "Gaussian processes in machine learning." Summer school on machine learning. Springer, Berlin, Heidelberg, 2003 a so-called Gaussian Processes known as a posteriori method, which can represent temporal, spatial or any other functions whose function values can only be modeled with certain uncertainties and probabilities due to incomplete information.

Advantages of the invention

The invention with the features of independent claim 1 allows, given a distribution of functions or heuristics, which roughly model the learning problem to be solved and which are all collectively referred to below as prior, a modeling of a posteriori distribution after a small number seen from exemplary data points. This posteriori modeling has several advantages: (i) It has a theoretical basis and is therefore correct when there is a situation that resembles one of the prior functions, (ii) it can therefore better take into account the actual probability of various events, it is consequently better "calibrated" and (iii) it enables particularly precise few-shot learning due to the advantage (i).

Further aspects of the invention are the subject of the independent claims. Advantageous further developments are the subject of the dependent claims.

Disclosure of the invention

In a first aspect, the invention relates to a computer-implemented method for teaching a machine learning system, in particular for predicting an output variable as a function of a plurality of input variables. At least one input of the machine learning system is suitable for receiving at least one pair of variables and processing them into an output variable for the machine learning system. A pair of sizes can be understood to mean that two, in particular different, sizes are assigned to one another and thus form a pair. The output variable of the machine learning system is preferably a vector for a classification or the like or a variable for a regression.

The machine learning system preferably has a plurality of inputs and one output and particularly preferably a temporal memory, the inputs being set up to each receive a pair of variables. The output is set up to output an output variable which is preferably a probability distribution or a continuous number. The temporal memory of the machine learning system can be given, for example, by recurrent connections between the layers or a buffer that provides or adds at least an intermediate result of a previous calculation of a previous input variable when processing a current input variable. The temporal memory is thus a form of providing information that was determined during an (immediately) preceding calculation of the machine learning system.

The machine learning system is, for example, a recurrent neural network, or the machine learning system can accommodate a plurality of pairs at the same time; a convolutional neural network can be conceivable here, the filters being set up in such a way that they generate the temporal memory by placing them over a row of the couples are applied. Attention-based networks are preferred, such as a transformer (Vaswani et al. / Https://arxiv.org/abs/1706.03762). The transformer's temporal memory is given by its 'attentions' and thus the transformer can look at every previous input / time step (or, depending on the mask, into the future) for each input or for each time step.

The method begins with providing a first probability distribution (also referred to as prior p (f)), which describes a probability distribution of a plurality of heuristics (f). The heuristics can be predetermined and are, in particular, functions which present simplified or presumptive conclusions about a system. Generally speaking, the heuristic assigns an output variable to an input variable. The input variable and the assigned output variable can form a pair. Alternatively, it is conceivable that the prior p (D, x, y) describes the pairs directly. The input variable is preferably a sensor variable that was recorded, for example, on a technical system. The output variable is preferably a parameter, in particular of the technical system or the sensor variable. The heuristics (f) can output estimates of a parameter to be assigned for a detected sensor variable in different ways or on the basis of different criteria.

1. a) Zufälliges Auswählen einer der Heuristiken (f) abhängig von der ersten Wahrscheinlichkeitsverteilung.
2. b) Erzeugen von Trainingsdaten (D) abhängig von der ausgewählten Heuristik (f), wobei einer Mehrzahl von Sensorgrößen mittels der Heuristik (f) jeweils eine Kenngröße zugeordnet wird. Die Sensorgrößen können reale, erfasste Größen mittels eines Sensors oder künstlich erzeugte Größen, z.B. mittels eines virtuellen Sensors, sein.
3. c) Anlernen des maschinellen Lernsystemes anhand der Trainingsdaten (D), wobei die Sensorgrößen und zugehörigen Kenngrößen als Paare dem maschinellen Lernsystem zugeführt werden, wobei das Anlernen derart erfolgt, dass der Ausgang des maschinellen Lernsystems im Wesentlichen die Kenngröße ausgibt. Unter im Wesentlichen kann hier eine zu erzielenden Genauigkeit, die mit dem jeweiligen vorliegenden Trainingsalgorithmus und Datensatz realisierbar ist, verstanden werden. Es sei angemerkt, dass für den Schritt des Anlernens jedes beliebiges Trainingsverfahren verwendet werden kann, z.B. bekannte Gradientenabstiegsverfahren.

The following steps a) to c) are then repeated several times. It should be noted that here the repetitions can be canceled, for example, after a specifiable maximum number of repetitions. However, the more repetitions that are performed, the better the machine learning system becomes.

1. a) Randomly selecting one of the heuristics (f) as a function of the first probability distribution.
2. b) Generation of training data (D) as a function of the selected heuristic (f), a parameter being assigned to a plurality of sensor variables by means of the heuristic (f). The sensor variables can be real, recorded variables by means of a sensor or artificially generated variables, for example by means of a virtual sensor.
3. c) Teaching the machine learning system on the basis of the training data (D), the sensor variables and associated parameters being fed to the machine learning system as pairs, the teaching being carried out in such a way that the output of the machine learning system essentially outputs the characteristic. Essentially, an accuracy to be achieved that can be achieved with the respective training algorithm and data set that is present can be understood here. It should be noted that any desired training method can be used for the learning step, for example known gradient descent methods.

A majority of the pairs are generated before the teaching. In a preferred embodiment, this plurality of pairs is divided into a first part of pairs and a second part. The second part is used as query points (querry points), at which only the sensor variable is then fed to the machine learning system, with the associated parameter of the respective pair expected at the output, with the machine learning system then also being adapted accordingly, that it outputs the assigned parameters according to the respective pairs.

One advantage of the method is that the heuristics can be selected flexibly and training data can be generated with them at low cost. Interestingly, very simple heuristics can be selected and the learned machine learning system can still perform well in an inference phase.

The learning step can include at least the following sub-steps: Calculating a cost function (l _θ ), the cost function being given to the pairs by a sum of a respective logarithm applied to the determined output variables (q _θ (y | x, D)) of the machine learning system is determined from the training data set (D) weighted with the probability (p (y, x, D)) of the respective pairs from the training data set (D), and adjusting parameters of the machine learning system in such a way that the cost function (l _θ ) is optimized , in particular by gradient descent methods, so that the cost function (l _θ ) is maximized or minimized by adapting parameters of the machine learning system.

wobei p(y,x,D) eine Wahrscheinlichkeit ist, welche charakterisiert, wie wahrscheinlich die vorliegende Kombination der Größen x und y aus Trainingsdaten D ist. Vorzugweise wird diese Wahrscheinlichkeit durch eine Unifromverteilung (auch als Gleichverteilung bezeichnet) modelliert. Der zweite Term der Kostenfunktion log q_θ(y|x,D) entspricht einem Logarithmus angewendet auf die Ausgabe des maschinellen Lernsystems, wobei dann die Ausgabe des maschinellen Lernsystems eine Wahrscheinlichkeitsverteilung beschreibt.The cost function l _θ can be defined formally: $$l_{\theta }={\displaystyle {\int }_{\mathrm{D.},x,y}p\left(y,x,\mathrm{D.}\right)\text{log}{q}_{\theta }\left(y|x,\mathrm{D.}\right)}$$

where p (y, x, D) is a probability which characterizes how probable the present combination of the quantities x and y from training data D is. This probability is preferably modeled by a uniform distribution (also referred to as uniform distribution). The second term of the cost function log q _θ (y | x, D) corresponds to a logarithm applied to the output of the machine learning system, the output of the machine learning system then describing a probability distribution.

wobei p(x,D) eine Wahrscheinlichkeit charakterisiert, dass der Datenpunkt x aus dem Trainingsdatensatz D gezogen wird. Vorzugweise wird diese Wahrscheinlichkeit durch eine Unifromverteilung modelliert. In anderen Worten kann gesagt werden, dass die Kostenfunktion eine cross-entropy über die Samples des Priors und den Vorhersagen des maschinellen Lernsystems ist.It is proposed that when calculating a cost function (l _θ ), the determined output variables (q _θ (y | x, D)) of the machine learning system with the probability p (x, D) that the respective sensor variable from the training data set (D ) is drawn, is weighted. Formally, the cost function (l _θ ) can be defined here: $$l_{\theta }={\displaystyle {\int }_{\mathrm{D.},x,y}p\left(y,x,d\right)\text{log}\left(q_{\theta }\left(y|x,\mathrm{D.}\right)p\left(x,\mathrm{D.}\right)\right)}$$

where p (x, D) characterizes a probability that the data point x will be drawn from the training data set D. This probability is preferably modeled by a uniform distribution. In In other words, it can be said that the cost function is a cross-entropy over the samples of the prior and the predictions of the machine learning system.

This slight manipulation of the cost function has the advantageous effect that the fine-grained learning takes into account the distribution of the sensor variables in the training data set, which is later expressed positively in a more precise prediction of the characteristic variable.

It is also proposed that the heuristics generate fictitious sensor variables based on random variables and output their assigned parameters. Fictitious sensor variables can be understood to mean that they were not recorded by a physical sensor, but rather simulated or calculated artificially.

The heuristics can represent assumptions about the problem at hand to be learned. For example, for a classification of the MNIST data set, the heuristic can generate simple black / white images with one or a plurality of randomly running straight lines, whereby the artificially generated images are classified as a function of the generated straight lines, for example a classification with regard to a number of the mapped ones Straight lines. Another heuristic of the prior can produce curved lines instead of straight lines, and yet another heuristic of the prior can be a combination of both of the aforementioned heuristics. The surprising result here is that after viewing a plurality of real MNIST images and classifying them, the machine learning system is able to classify the MNIST images, although the machine learning system has not been trained on these images.

The advantage here is that the machine learning system is taught without real training data, since real training data are usually complex to create.

A second aspect of the invention relates to a method for operating the machine learning system, which was learned, for example, according to the first aspect. The method of the second aspect begins with the provision of at least one pair of a sensor variable and assigned characteristic variable and a further sensor variable whose characteristic variable is to be predicted. The pairs represent samples that are, for example, recorded on a random basis when operating a technical system; in particular, their associated parameters are also recorded when operating the technical system. These pairs are preferably real sensor variables. The associated parameters can also be real, or they can be added manually to the sensor variables, also referred to below as labels. The advantage here is that a prediction can be made from one data pair onwards.

This is followed by iterative processing of the majority of the pairs by the machine learning system. Iterative processing can be understood to mean that the input variables and assigned output variables of the pairs are fed to the inputs of the machine learning system, with one pair being processed and / or one output variable being output for each iteration. However, iterative processing can also be understood to mean that the pairs are fed to the machine learning system at the same time and the machine learning system, for example, iteratively or sequentially outputs an output variable for the individual pairs. It should be noted that the order of the iterative processing of the pairs is not relevant, i.e. that either the following pair or any other pair can be processed per processing step.

It is conceivable that the processing of the data pairs can be understood as a form of re-learning the machine learning system. A special feature of this re-learning is that the parameters of the machine learning system no longer have to be optimized here. Because experiments have surprisingly shown that the machine learning system is able to 'select' the correct prior in order to estimate the posteriori. Optionally, however, the parameters can be optimized in the step of so-called post-learning, so that the outputs of the machine learning system for the pairs provided correspond to the output variables of the pairs.

After retraining, the processing of the individual input variable, the output variable of which is to be predicted, is carried out by the machine learning system. This step can also be referred to as an inference.

It is conceivable that the inference step is repeated several times, e.g. with the same or with different data points.

Accordingly, the pairs are iteratively processed by the machine learning system and then the last input variable is processed by the machine learning system, the machine learning system then predicting the output variable of the last input variable.

One advantage of the method of the second aspect of the invention is that, due to the special operation of the learned machine learning system, it is able to do so with very few Pairs of input variables to determine the output variable. It is therefore particularly data efficient.

The input variables of the first and second aspects are preferably sensor variables of a technical system and the assigned or determined output variables are parameters of the technical system.

It should also be noted that the machine learning system of the first and second aspects is set up so that, when it predicts the output variable, it determines a distribution over a predetermined set of output variables.

For the first and second aspect of the invention it is proposed that the machine learning system is a transformer, the input of the transformer being set up to receive a plurality of pairs. Neural networks as transformers are known, for example, from the disclosure: Ashish Vaswani et al: "Attention is all you need". In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017 . URL https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf. It should be noted that in general any neural network is conceivable for the first and second aspect of the invention which has a temporal memory, for example a recurrent neural network. However, it has been found that transformers produce particularly good results.

An attention mask of the transformer is preferably constructed in such a way that attention is given from each pair to every other pair, as well as attention from the last sensor variable, the characteristic variable of which is to be predicted, to the preceding pairs. In other words, the transformer's “attentions” are adjusted in such a way that the representation of each entry in the data set can depend on each other and the representation of the queried input can depend on itself and on the representations of the data set.

• i) Beide Größen der Paare werden jeweils durch einen Encoder komprimiert, der vorzugsweise eine affine Projektion ausführt, wobei beide komprimierten Größen der Paare zusammenaddiert werden und die Summe durch den Transformer verarbeitet wird, und
• ii) Vorzugweise werden keine „Positional Encodings“ verwendet.

To make the transformer invariant to the order of the pairs, it is suggested:

• i) Both sizes of the pairs are each compressed by an encoder, which preferably carries out an affine projection, the two compressed sizes of the pairs being added together and the sum being processed by the transformer, and
• ii) Preferably no “positional encodings” are used.

It is also proposed that the output variables of the machine learning system are discrete and that the machine learning system assigns probabilities to the possible output variables, with a smallest and largest output variable being assigned to a correspondingly scaled half of a normal distribution.

This has the advantage that a possible representation of a continuous multimodal probability distribution by so-called "bars" is possible, and the scaled normal distribution at both ends also gives a probability outside the specified values. This prevents a cost function (loss) from increasing to infinity, especially for regressions in which very large values occur. In other words, based on this approach, the machine learning system can also be used as a regressor.

It is further proposed that the learned machine learning system according to the first aspect of the invention is used to classify sensor signals according to the second aspect. The output variables of the machine learning system then characterize a classification of the sensor variables. The sensor variables are preferably images and the output variables are at least classifications of the images. The machine learning system is then an image classifier. The image classifier assigns an input image to one or more classes of a predetermined classification. For example, images of serially manufactured, nominally identical products can be used as input images. The image classifier can, for example, be trained to assign the input images to one or more of at least two possible classes that represent a quality assessment of the respective product. The concept of the image basically encompasses any distribution of information arranged in a two- or multi-dimensional grid. This information can be, for example, intensity values of image pixels that were recorded with any desired imaging modality, such as with an optical camera, with a thermal imaging camera or with ultrasound. However, any other data, such as audio data, radar data or LIDAR data, can also be translated into images and then classified in the same way.

It is also conceivable, however, that the sensor variables and preferably also the output variables are measurements, in particular of a metallic component during a heat treatment. That is to say, the sensor variables can also be recorded system states of a technical system that were recorded, for example, by means of a sensor. And the output variables are, for example, parameters and / or subsequent states of the technical system, electrical or mechanical component. For example, the output variables are values of a characteristic curve or a characteristic diagram.

It is further proposed that the machine learning system, which was operated according to the second aspect, determines an output variable as a function of a detected sensor variable of a sensor, depending on which a control variable is then determined by means of, for example, a control unit.

The control variable can be used to control an actuator of a technical system. The technical system can be, for example, an at least partially autonomous machine, an at least partially autonomous vehicle, a robot, a tool, a work machine or a flying object such as a drone. The input variable can, for example, be determined as a function of the recorded sensor data and made available to the machine learning system. The sensor data can be recorded by a sensor, such as a camera, of the technical system or, alternatively, can be received externally.

In further aspects, the invention relates to a device and a computer program, each of which is set up to carry out the above methods, and a machine-readable storage medium on which this computer program is stored.

• 1 schematisch ein Flussdiagramm einer Ausführungsform der Erfindung;
• 2 schematisch ein Mehrzahl von Funktionen und Stichproben einer der Funktionen sowie exemplarische Darstellungen von Ausgaben eines maschinellen Lernsystems;
• 3 schematisch ein Ausführungsbeispiel zur Steuerung eines wenigstens teilautonomen Roboters;
• 4 schematisch ein Ausführungsbeispiel zur Steuerung eines Fertigungssystems;
• 5 schematisch ein Ausführungsbeispiel zur Steuerung eines Zugangssystems;
• 6 schematisch ein Ausführungsbeispiel zur Steuerung eines Überwachungssystems;
• 7 schematisch ein Ausführungsbeispiel zur Steuerung eines persönlichen Assistenten;
• 8 schematisch ein Ausführungsbeispiel zur Steuerung eines medizinisch bildgebenden Systems;
• 9 schematisch eine Trainingsvorrichtung.

Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. In the drawings show:

• 1 schematically, a flow diagram of an embodiment of the invention;
• 2 schematically a plurality of functions and samples of one of the functions as well as exemplary representations of outputs of a machine learning system;
• 3 schematically an embodiment for controlling an at least partially autonomous robot;
• 4th schematically an embodiment for controlling a manufacturing system;
• 5 schematically an embodiment for controlling an access system;
• 6th schematically an embodiment for controlling a monitoring system;
• 7th schematically an embodiment for controlling a personal assistant;
• 8th schematically an embodiment for controlling a medical imaging system;
• 9 schematically a training device.

Description of the exemplary embodiments

1 shows a flowchart of a method for teaching a machine learning system so that the machine learning system that has been taught can be used for “fewshot learning”. The term “few-shot learning” refers to a learning paradigm in which only a very small amount of training data is available, which is in contrast to the large amounts of data commonly used for deep learning, for example. Also shows 1 optional downstream steps to operate the learned machine learning system for “few-shot learning”.

The procedure begins with step S11 , providing the machine learning system and a prior p (ƒ), which describes a probability distribution of a plurality of functions /, the functions / outputting which output variable is assigned to the input variables. The functions can implement heuristics that output a probable statement about the input variables.

The machine learning system is suitable for receiving input / inputs as a function of an input variable and an output variable assigned to the input variable and processing it into an output variable of the machine learning system. The machine learning system preferably outputs a probability distribution over the possible output variables. This is exemplified in 2 by means of the representations 200-202 shown as an example.

In a preferred embodiment of the method according to 1 the machine learning system is given by a transformer. It is known that transformers differ from (deep) neural networks in that transformers have at least one additional layer that perform “sef-attention” when the information is propagated through the transformer.

An attention mask of the transformer describes which inputs are taken into account in relation to each other. The attention mask is preferably constructed in such a way that the attention of all processed pairs to one another and an attention of the last sensor variable to all processed pairs to one another characterizes. The attention mask M _{m, n} can be given, for example, for m pairs of input and output variables of the transformer and n query points for e.g. 'few-shot learning', here m, n = 4.2: $${\mathrm{M.}}_{4.2}=\left[\begin{array}{llllll}1\hfill & 1\hfill & 1\hfill & 1\hfill & 0\hfill & 0\hfill \\ 1\hfill & 1\hfill & 1\hfill & 1\hfill & 0\hfill & 0\hfill \\ 1\hfill & 1\hfill & 1\hfill & 1\hfill & 0\hfill & 0\hfill \\ 1\hfill & 1\hfill & 1\hfill & 1\hfill & 0\hfill & 0\hfill \\ 1\hfill & 1\hfill & 1\hfill & 1\hfill & 1\hfill & 0\hfill \\ 1\hfill & 1\hfill & 1\hfill & 1\hfill & 0\hfill & 1\hfill \end{array}\right]$$

gezogen.The forecast of the transformer can be output after the last processing step of the transformer. A total size M = m + n and a sample set m are preferably established during the training so that the transformer is able to work with different large data sets. Since there are more query points (querry points) for smaller data sets, more frequent pulling (over-sample) can be carried out for larger data sets. That means, every me {0, ... M - 1} becomes dependent on a weight $\frac{1}{n}=\frac{1}{\mathrm{M.}-m}$

drawn.

• In Schritt S12 wird ein zufälliges Auswählen einer der Funktionen / aus dem Prior p(ƒ) durchgeführt.

After step S11 the following steps are repeated several times S12 until S14 :

• In step S12 a random selection of one of the functions / from the prior p (ƒ) is carried out.

After step S12 follow step S13 . This is where training data D is generated depending on the function / from step S12 executed. This is carried out, for example, in that a plurality of randomly generated input variables x by means of the function / from step S12 an output variable y is assigned to each of the input variables. This is exemplified in 2 through the plot 20th illustrates in which the functions 21a , 21b assign a y-value to the x-values.

After generating the training data D after step S13 will step S14 executed. The machine learning system is taught in here on the basis of the training data D. The training data are propagated by the machine learning system, a cost function being determined depending on the output variables determined by the machine learning system and the output variables of the training data D, which is then optimized by a gradient descent method with regard to the parameters of the machine learning system. The cost function can be a cross-entropy between the output variables of the training data D and the output variables determined by the machine learning system. The cost function is preferably determined for a batch of training data D, with the step S14 running for a variety of batches.

When step S14 finished, the steps can be repeated S12 until S14 be carried out again or the method for teaching the machine learning system is ended if a termination criterion has been met.

Optionally, after completion of the step S14 the learned machine learning system can be operated in accordance with the steps explained below. This is done in step S15 a plurality of pairs of input variables and assigned output variables and a last input variable whose output variable is to be predicted is provided.

Then the next step S16 iteratively the pairs from step S15 processed by the machine learning system. In other words, the pairs are fed to the inputs of the machine learning system and propagated through it. It should be noted that the output variables of the machine learning system determined on this basis are not relevant and can be discarded.

Then step S17 the last input variable, which does not have an assigned output variable, is fed to one of the inputs, in particular to that input whose dimensionality is compatible with the last input variable. The machine learning system then outputs a prediction of a possible assigned output variable. A representation 200-202 the prediction is in 2 shown as an example.

After step S17 has ended, the procedure for operating is ended. As new pairs are deployed, the steps can S15 until S17 to be run again. If a further input variable, which does not have an assigned output variable, is provided, only step can also be used S17 are executed.

2 shows schematically a representation 200-202 a prediction of the machine learning system and a plot 20th a variety of functions 21b as well as a highlighted function 21a . The functions 21a , 21b are described by the prior p.

The first representation 200 a prediction y _{pred of} the machine learning system shows, for example, that the machine learning system assigns values between -3 and 3 to the input variables, the value range from -3 to 3 being divided into intervals and each interval being assigned a probability (prob.). The first representation 200 is characterized in that the intervals are essentially the same. Conveniently, the machine learning system for this representation 200 output a vector whose Entries are each assigned to an interval and the entries are the probabilities (prob.).

The second representation 201 differs from the first illustration 201 in that the individual intervals are of different widths from one another.

The third representation 202 differs from the first and second representation 202 essentially that the two outermost intervals, which cover the two ends of the value range, have been replaced by scaled, halved normal distributions.

plot 20th schematically shows the operation of the machine learning system according to FIG 1 . Spot checks 22nd are different from that of the highlighted function 21a determined, taking the samples 22nd are each defined by an x and y value. These samples are then taken according to step S16 processed by the machine learning system. Based on the samples 23 which has only one x _q value, the corresponding y value is to be predicted. For this purpose, the sample 23 according to step S17 processed by the machine learning system, the machine learning system then outputting the associated y value, or according to one of the representations 200-202 .

In a preferred embodiment of the invention, the method is according to 1 used for a heat treatment process for hardening metallic materials. For example, in inductive heat treatment, an alternating magnetic field is generated with the aid of an inductor. The alternating magnetic field induces eddy currents in the workpiece to be hardened, which locally lead to an increase in the temperature in the workpiece. The induction hardening process enables component properties, such as the hardening depth, SHD and the like, to be adjusted. On the one hand, variations in the alloy composition and different structural characteristics of the starting material of workpieces due to the resulting distribution of the local electrical resistance lead to fluctuations and variations in the process temperatures and thus to sometimes considerable variations in the final component properties. On the other hand, if the material, the component geometry and / or the processing tools are changed, extensive re-parameterization of the heat treatment process is necessary.

To parameterize a heat treatment process, extensive input tests are carried out for each component geometry and material in order to determine a stable process window for the process parameters of the heat treatment process. Experience from previous measurements of other workpiece geometries and materials can only be partially taken into account. This increases the time required to parameterize the heat treatment process for a component type, since a single measurement of a workpiece is very complex. It is known that, in order to determine process parameters for a heat treatment process of a component, in particular for hardening the component, the process parameters are determined with the aid of a Bayesian optimization method.

It is now proposed that instead of a Gaussian process model for Bayesian optimization for predicting the quality of the treated components, the learned machine learning system should be used 1 is used. A prior can be used to train the machine learning system, which maps the input variables to quality measures in the form of physical heuristics. The prior can thus specify a quality of a component in terms of one or more component properties and / or process properties.

After measuring components that are subjected to a heat treatment process in accordance with test process parameters with regard to the one or more component properties, the Gaussian process model does not have to be optimized, because according to the steps S15 until S17 after 1 it is sufficient if the measurements are propagated by the learned machine learning system.

According to Bayesian optimization, an acquisition function is created in order to obtain a next test process parameter set for measuring the one or more component properties of a correspondingly manufactured component. The acquisition function can then be output after step S17 processed by the machine learning system in order to predict whether the selected one of the process parameters lead to an improved quality. The acquisition function is implemented using the machine learning system. The acquisition function can be freely selected here. In particular, however, it should be explained that the very popular acquisition function Expected Improvement is very easy to implement with the help of the bar distribution, as in 2 and ultimately simplified to a simple sum across the bars.

A heat treatment process generally provides for a heat treatment of a workpiece over a total period of time, the temperature profile following a predetermined trajectory during the entire period of time. The temperature trajectory is chosen so that certain component properties are to be achieved. The parameterization the process parameter for setting the temperature trajectory takes place according to a result of a Bayesian optimization method, which is based on a trainable quality function in the form of a Gaussian process. The quality function maps the process parameters of the heat treatment process to a relevant quality measure, which characterizes component parameters such as the structural composition and / or process properties (such as the process duration) of the manufacturing process. In addition, requirements for the process result, such as a minimum hardness of the component to be hardened, can be taken into account as secondary conditions. These secondary conditions are taken into account in the optimization process with further Gaussian process models.

• - eine Zeitdauer für eine Bearbeitungsphase;
• - einen Zeitpunkt des Wechselns zwischen Bearbeitungsphasen;
• - eine Bearbeitungstemperatur in einer Bearbeitungsphase; und
• - eine Leistungszufuhr zum Erwärmen des Bauteils in einer Bearbeitungsphase.

Furthermore, the heat treatment process can be divided into several successive processing phases, the process parameters specifying one or more of the following parameters:

• - a period of time for a processing phase;
• - a point in time of switching between processing phases;
• a machining temperature in a machining phase; and
• - A power supply for heating the component in a processing phase.

In a preferred embodiment of the invention, the method is according to 1 used for a hyperparameter optimization when teaching another machine learning system. Here, the input variables can be the hyperparameters such as a learning rate or the like, for which a performance of the additional machine learning system is then predicted with these hyperparameters.

In a further embodiment of the method according to 1 the machine learning system is used for an actuator control system.

3 shows schematically an actuator with a control system 40 . An environment is created at preferably regular time intervals 20th of the actuator 10 with a sensor 30th , in particular an imaging sensor such as a video sensor, which can also be provided by a plurality of sensors, for example a stereo camera. Other imaging sensors are also conceivable, such as, for example, radar, ultrasound or lidar. A thermal imaging camera is also conceivable. The sensor signal S - or, in the case of several sensors, one sensor signal S - of the sensor 30th is attached to the control system 40 transmitted. The control system 40 thus receives a sequence of sensor signals S. The control system 40 determines control signals A from this, which are sent to an actuator 10 be transmitted. The actuator 10 can convert received control commands into mechanical movements or changes in physical quantities. The actuator 10 can e.g. B. convert the control command A into an electrical, hydraulic, pneumatic, thermal, magnetic and / or mechanical movement or cause a change. Specific but non-limiting examples are electric motors, electroactive polymers, hydraulic cylinders, piezoelectric actuators, pneumatic actuators, servomechanisms, solenoids, stepper motors, etc.

The control system 40 receives the sequence of sensor signals S from the sensor 30th in an optional receiving unit 50 , which converts the sequence of sensor signals S into a sequence of input images x (alternatively, the sensor signal S can also be directly adopted as the input image x). The input image x can be a section or further processing of the sensor signal S, for example. The input image x comprises individual frames of a video recording. In other words, the input image x is determined as a function of the sensor signal S. The sequence of input images is therefore x 1 learned machine learning system 60 , fed.

According to the steps S15-S17 determines the machine learning system 60 an output variable y from the input images x. These output variables y can in particular include a classification and / or semantic segmentation of the input images x. Output variables y are an optional forming unit 80 supplied, which determines from this control signals A, which the actuator 10 be fed to the actuator 10 to be controlled accordingly. Output variable y includes information about objects that the sensor 30th has captured.

The actuator 10 receives the control signals A, is controlled accordingly and carries out a corresponding action. The actuator 10 can in this case comprise a control logic (not necessarily structurally integrated) which determines a second control signal from the control signal A, with which the actuator is then 10 is controlled.

In further embodiments, the control system comprises 40 the sensor 30th . In still further embodiments, the control system comprises 40 alternatively or additionally also the actuator 10 .

In further preferred embodiments, the control system comprises 40 a single or multiple processors 45 and at least one machine-readable storage medium 46 , on which instructions are stored, which then when they are on the processors 45 run that Control system 40 cause to carry out the method according to the invention.

In alternative embodiments it is an alternative or in addition to the actuator 10 a display unit 10a provided, which is an output variable of the control system 40 can display.

In other embodiments, the display unit 10a be an output interface to a playback device, such as e.g. B. a display, a light source, a loudspeaker, a vibration motor, etc., which can be used to generate a sensory perceptible output signal based on the output of the machine learning system 60 can be generated.

In a preferred embodiment of the 2 becomes the control system 40 to control the actuator, which here is an at least partially autonomous robot, here an at least partially autonomous motor vehicle 100 , is, used. With the sensor 30th it can be, for example, one preferably in the motor vehicle 100 act arranged video sensor.

Preferably in the motor vehicle 100 arranged actuator 10 it can for example be a brake, a drive or a steering of the motor vehicle 100 Act. The control signal A can then be determined in such a way that the actuator or actuators 10 is controlled in such a way that the motor vehicle 100 for example a collision with the machine learning system 60 safely identified objects are prevented, especially if they are objects of certain classes, for example pedestrians.

Alternatively, the at least partially autonomous robot can also be another mobile robot (not shown), for example one that moves by flying, swimming, diving or walking. The mobile robot can also be, for example, an at least partially autonomous lawnmower or an at least partially autonomous cleaning robot. In these cases too, the control signal A can be determined in such a way that the drive and / or steering of the mobile robot are controlled in such a way that the at least partially autonomous robot, for example, has a collision with the artificial neural network 60 identified objects.

4th shows an embodiment in which the control system 40 for controlling a production machine 11 of a manufacturing system 200 is used by a this manufacturing machine 11 controlling actuator 10 is controlled. At the manufacturing machine 11 it can be, for example, a machine for punching, sawing, drilling, milling and / or cutting.

With the sensor 30th it can then be, for example, an optical sensor which, for example, measures the properties of manufactured products 12a , 12b recorded. It is possible that these manufactured articles 12a , 12b are movable. It is possible that the the manufacturing machine 11 controlling actuator 10 depending on an assignment of the recorded production products 12a , 12b is controlled so that the production machine 11 corresponding to a subsequent processing step of the correct one of the manufactured product 12a , 12b executes. It is also possible that by identifying the correct properties of the same of the manufactured products 12a , 12b (ie without a misallocation) the manufacturing machine 11 accordingly adjusts the same production step for processing a subsequent production product.

5 shows an embodiment in which the control system 40 to control an access system 300 is used. The access system 300 can be a physical access control, for example a door 401 include. Video sensor 30th is set up to capture a person. Using the object identification system 60 this captured image can be interpreted. If several people are recorded at the same time, the identity of the people can be determined particularly reliably by assigning the people (that is to say the objects) to one another, for example by analyzing their movements. The actuator 10 can be a lock that releases the access control depending on the control signal A, or not, for example the door 401 opens or not. For this purpose, the control signal A can be dependent on the interpretation of the object identification system 60 be selected, for example depending on the identified identity of the person. Instead of the physical access control, a logical access control can also be provided.

6th shows an embodiment in which the control system 40 for controlling a monitoring system 400 is used. From the in 5 The illustrated embodiment differs from this embodiment in that instead of the actuator 10 the display unit 10a is provided by the control system 40 is controlled. For example, from the machine learning system 60 reliably identifies the identity of the video sensor 30th recorded objects are determined in order to infer, for example, as a function of this which are suspect, and the control signal A is then selected in such a way that this object is shown by the display unit 10a highlighted in color.

7th shows an embodiment in which the control system 40 to control a personal assistant 250 is used. The sensor 30th is preferably an optical sensor, the images of a gesture by a user 249 receives.

Depending on the signals from the sensor 30th determines the control system 40 a control signal A of the personal assistant 250 , for example, by the neural network performing gesture recognition. The personal assistant 250 this determined control signal A is then transmitted and it is thus controlled accordingly. This determined control signal A act can in particular be selected in such a way that it is a presumed desired control by the user 249 is equivalent to. This presumed desired control can depend on the artificial neural network 60 recognized gesture can be determined. The control system 40 can then, depending on the presumed desired control, the control signal A for transmission to the personal assistant 250 and / or the control signal A for transmission to the personal assistant according to the presumed desired control 250 Select.

This corresponding control can include, for example, that the personal assistant 250 Retrieves information from a database and makes it available to the user 249 reproduces in a receivable manner.

Instead of the personal assistant 250 A household appliance (not shown), in particular a washing machine, a stove, an oven, a microwave or a dishwasher, can also be provided in order to be controlled accordingly.

8th shows an embodiment in which the control system 40 for controlling a medical imaging system 500 , for example an MRI, X-ray or ultrasound machine is used. The sensor 30th can for example be given by an imaging sensor, by the control system 40 becomes the display unit 10a controlled. For example, from the neural network 60 it can be determined whether an area recorded by the imaging sensor is conspicuous, and the control signal A can then be selected in such a way that this area is detected by the display unit 10a highlighted in color.

9 shows schematically a training device 500 includes a provider 51 , which is generated from the training data set according to step S13 provides the training data. These training data are used by the machine learning system 52 supplied, which determines output variables therefrom. Output variables and input images become an assessor 53 supplied, the parameter updated from this after step S14 determined, which are transmitted to the parameter memory P and replace the current parameters there.

The one from the training device 500 The methods carried out can be implemented as a computer program on a machine-readable storage medium 54 be stored and by a processor 55 are executed.

The term “computer” encompasses any device for processing specifiable arithmetic rules. These calculation rules can be in the form of software, or in the form of hardware, or also in a mixed form of software and hardware.

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.

Non-patent literature cited

• Ashish Vaswani et al: "Attention is all you need". In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017 [0029]

Claims (11)

Apparatus for teaching a machine learning system (60), wherein an input of the machine learning system (60) is suitable for receiving at least one pair of variables and _{processing them into an output variable (y pred} ) of the machine learning system (60), the apparatus being a machine-readable one Storage medium comprises, on which instructions are stored which, when executed by a computer, caused the computer to carry out a method with the following steps: providing (S11) a first probability distribution (p (f)) which contains a probability distribution of a plurality of heuristics (f ) describes, the heuristics (f) outputting estimates of a parameter for a sensor variable; Repeating the steps several times: randomly selecting (S12) one of the heuristics (f) as a function of the first probability distribution (p (f)); Generation (S13) of training data (D) as a function of the selected heuristic (f), a parameter being assigned to a plurality of sensor variables by means of the heuristic (f); and teaching (S14) the machine learning system (60) on the basis of the training data (D), the sensor variables and associated parameters being fed as pairs to the machine learning system (60). Device according to Claim 1 , the stored commands being designed in such a way that the method that the computer executes when these commands are executed on the computer runs in such a way that the learning step (S14) comprises at least the following substeps: Calculating a cost function (l _θ ) , where the cost function is a sum of an applied logarithm to the determined output variables (q _θ (y | x, D)) of the machine learning system (60) given the pairs from the training data set (D) weighted with the probability (p (y , x, D)) of the respective pairs is determined from the training data set (D), and adapting parameters of the machine learning system in such a way that the cost function (l _θ ) is optimized. Device according to Claim 2 , the stored commands being designed in such a way that the method that the computer executes when these commands are executed on the computer runs in such a way that when calculating a cost function (l _θ ), the determined output variables (q _θ (y | x , D)) of the machine learning system (60) is weighted with a probability (p (x, D)) that the respective sensor variable (x) is drawn from the training data set (D). Device according to one of the preceding claims, wherein the stored commands are designed such that the method that the computer executes when these commands are executed on the computer runs in such a way that the heuristics based on random variables generate fictitious sensor variables and output their assigned characteristic variables . Device for operating the machine learning system (60), which has been trained according to one of the preceding claims, wherein the device comprises a machine-readable storage medium on which instructions are stored which, when executed by a computer, caused the computer to execute a method with the following steps : Providing (S15) a pair or a plurality of pairs of sensor variables and associated parameters and a further sensor variable whose characteristic is to be predicted; Processing (S16) the pair or the plurality of pairs by the machine learning system (60); Processing (S17) the further sensor variable by the machine learning system (60) and outputting the determined parameter by the machine learning system. Device according to one of the preceding claims, wherein the stored commands are designed such that the method that the computer executes when these commands are executed on the computer runs in such a way that the machine learning system (60) is a transformer, the input of the transformer is arranged to receive a plurality of pairs. Device according to Claim 6 The stored commands are designed in such a way that the method that the computer executes when these commands are executed on the computer runs in such a way that a mask of the transformer processed an attention of all processed pairs to each other and an attention of the further sensor variable to all Couples characterized to one another. Device according to Claim 6 or 7th , wherein the stored commands are designed such that the method that the Computer executes, when these commands are executed on the computer, the sequence is that both sizes of the pairs are compressed by an encoder, the two compressed sizes of the pairs being added together and the sum being processed by the transformer. Device according to one of the preceding claims, wherein the stored commands are designed such that the method that the computer executes when these commands are executed on the computer runs in such a way that the output variables of the machine learning system are discrete and the machine learning system is the possible Assigns output variables to probabilities, with the smallest and largest possible output variable being assigned to a scaled half of a normal distribution. Computer program product comprising program code, which is stored on a computer-readable medium, comprising: Program code for providing (S11) a first probability distribution (p (f)) which describes a probability distribution of a plurality of heuristics (f), the heuristics (f) outputting estimates of a characteristic variable for a sensor variable; Program code for repeating the steps multiple times: Program code for randomly selecting (S12) one of the heuristics (ƒ) as a function of the first probability distribution (p (f)); Program code for generating (S13) training data (D) as a function of the selected heuristic (f), a characteristic variable being assigned to a plurality of sensor variables by means of the heuristic (f); and Program code for teaching (S14) the machine learning system (60) on the basis of the training data (D), the sensor variables and associated parameters being fed as pairs to the machine learning system (60). Computer program product according to Claim 10 wherein the program code additionally or alternatively comprises: program code for providing (S15) a pair or a plurality of pairs of sensor variables and associated parameters and a further sensor variable whose characteristic is to be predicted; Program code for processing (S16) the pair or the plurality of pairs by the machine learning system (60); Program code for processing (S17) the further sensor variable by the machine learning system (60) and outputting the determined parameter by the machine learning system.

Images