DE102019207911A1 - Method, device and computer program for predicting a learning curve - Google Patents

Abstract

The invention relates to a method (30) for predicting a learning curve of an iterative learning process. The method (30) comprises the steps of obtaining the hyperparameters θ of the training method and at least one previous value of the cost function and the step of predicting, by means of a second machine learning system (12), the learning curve depending on the hyperparameter θ and depending on the at least one previous value of the learning curve. The invention also relates to a computer program and a device for carrying out the method (30) and a machine-readable memory element on which the computer program is stored.

Description

Technical area

The invention relates to a method for predicting a learning curve in an iterative learning method of a machine learning system. The invention also relates to a device and a computer program which are each set up to carry out the method.

State of the art

The authors T. Elsken, J. Metzen, and F. Hutter give in their publication "Neural architecture search: A survey." ArXiv preprint arXiv: 1808.05377 (2018) an overview of methods for optimizing the architecture of machine learning systems.

The authors A. Klein, S. Falkner, J. T. Springenberg, and F. Hutter disclose in their publication "Learning curve prediction with Bayesian neural networks" International Conference on Learning Representations (ICLR'17) a method for predicting learning curves of machine learning systems.

Advantages of the invention

Teaching machine learning systems is very computationally intensive and therefore extremely time-consuming even on powerful computers. This problem is exacerbated when the machine learning system has many parameters (e.g. <100 to> 1 million parameters) and a lot of training data is required for teaching, as these have to be processed by the machine learning system during teaching. A hyperparameter of the learning process can be adjusted by trial and error in order to positively influence the learning curve. Trying it out is extremely inefficient, however, since all calculations have to be carried out every time without being able to foresee whether this adapted hyperparameter of the teaching process will subsequently lead to a better learning curve.

The aim of the invention is to simplify the learning process of machine learning systems by specifically determining a suitable value of the hyperparameters of the learning process, so that overall fewer learning steps are required, whereby the learning process becomes more efficient in terms of computer resources.

It is suggested that the learning curve be predicted inexpensively. Depending on the prediction, a decision can then be made as to whether the teaching with the selected hyperparameter of the teaching method will be expedient, that is, will be retained, or whether the hyperparameter should be changed. The hyperparameter of the learning method can then be used selectively when teaching the machine learning system in order to train the machine learning system more quickly and thus more economically with computer resources, without using unnecessarily many, non-targeting values of the hyperparameters, in particular testing them.

Furthermore, the invention also allows a hyperparameter of the learning method to be found, which makes it possible to find an optimum of the cost function during the learning process more precisely and thus to achieve a higher efficiency of the machine learning system.

Disclosure of the invention

In a first aspect, a, in particular computer-implemented, method according to independent claim 1 for predicting a learning curve of an iterative learning process of a first machine learning system is presented. The learning curve is a, in particular sequential, time series of a plurality of values which each correspond to a value of a cost function of the iterative learning method or which each characterize a performance of the first machine learning system on training data. Training data is understood to mean all the data provided, which each comprise a pair of input variables and assigned output variables and are used for teaching and evaluating the first machine learning system. A validation data set is preferably used to evaluate the performance.

The values of the learning curve are preferably each assigned to an iteration step of the learning method. The learning curve describes a course or a progress of the learning process, that is, a development over time of the values of the cost function or the performance along the iteration steps of the learning of the first machine learning system. The performance can be understood to mean the extent to which the first machine learning system, applied to the training data, is already performing its task, according to which the first machine learning system is trained to solve this task. The performance is, for example, a classification accuracy of the training data that is used to teach the first machine learning system. The iterative teaching method is used to train the first machine learning system and is parameterized by a parameterization. The method for predicting the learning curve comprises the following steps: Obtaining the parameterization of the iterative learning process and at least one previous value of the learning curve, which is based on at least one previously executed, preferably most current, Iteration step of the learning process was determined. In other words, the previous value of the learning curve essentially corresponds to a previous value of the cost function of this iteration step or characterizes the performance after this iteration step has been carried out. This is followed by a prediction, by means of a second machine learning system, of a subsequent value of the learning curve depending on the parameterization and depending on the at least one previous value of the learning curve. That is to say, the subsequent value of the learning curve essentially corresponds to a value of the cost function that the cost function would determine for a subsequent iteration step of the learning method or characterizes the performance after the subsequent iteration step would have been carried out. This is followed by a decision as a function of the predicted value of the learning curve as to whether the parameterization of the iterative learning method is retained. It is also conceivable that the steps of the method are repeated a number of times and, depending on a plurality of predicted values of the learning curve, a decision is made as to whether the parameterization of the iterative learning method is retained.

The advantage of the method is that it predicts the course of the learning curve particularly efficiently and reliably. Since the parameterization of the iterative learning process influences the learning curve and thus the success of the learning of the first machine learning system, this process can be used to predict how promising the parameterization will be. The parameterization can be understood to mean one or a plurality of hyperparameters. A hyperparameter is a parameter whose value (s) is selected at the beginning of a learning process and remains unchanged until the iterative learning process is ended.

It should be noted that the cost function can be evaluated for each iteration step of the learning method and a parameterization of the first machine learning system is adapted depending on the cost function. The parameterization of the first machine learning system can include more than 100 parameters or more than 1 million parameters.

An iterative learning process is understood to mean a (training) process that adapts the parameterization of the first machine learning system in such a way that the cost function becomes optimal with regard to a specifiable criterion. The parameterization of the first machine learning system is adapted iteratively. The parameterization of the first machine learning system is alternately adapted depending on the cost function and the cost function is then determined depending on the adapted parameterization. I.e. the parameterization is adapted, whereupon the cost function is updated and the parameterization is adapted again depending on the updated cost function. This can be done several times in succession until a predefinable termination criterion is reached. For example, the iterative learning process can be a gradient descent process such as Adam. The first machine learning system can be any type of adaptive system, such as a neural network or a Gaussian process or a support vector machine, etc. The first machine learning system can be used for the classification or segmentation of images, video sequences and / or audio signals. It should also be noted that the machine learning system can also be used for other tasks, such as a regression, can be used.

A cost function can be a difference or a deviation between at least one determined output variable of the first machine learning system as a function of a training input variable and a training output variable that is assigned to the training input variable. The cost function can for example be a cross entropy or the like.

It is suggested that the second machine learning system is a random forest. The random forest predicts the subsequent value of the learning curve as a function of a plurality of previous values of the learning curve. Each decision tree of the random forest outputs at least a first size and a second size. The prediction of the subsequent value of the learning curve is determined as a function of an average value over the first variables and as a function of an average value over the second variables.

The mean value of the first and second variables can be used to parameterize a probability distribution, preferably a Gaussian distribution, the predicted value of the learning curve then being randomly drawn from this probability distribution.

A random forest is a regression / classification method that consists of several uncorrelated decision trees. All decision trees were created during a learning process. For a classification, each tree in this "forest" can make a decision. The final classification is then made based on all decisions made by the trees. The first and second variables of the random forest preferably characterize a first and second moment of a probability distribution, for example a Gaussian distribution. The advantage of the Random Forest is that this type of machine learning system has cope with little training data and can therefore be efficiently created and operated.

It is further proposed that the second machine learning system is a Bayesian recurrent neural network, also known under the English name Variational Recurrent Neural Network (VRNN). It is further proposed that the Bayesian recurrent neural network comprises a plurality of LongShortTermMemory modules (LSTM) which are connected to one another in a predetermined sequence.

The advantage of recurrent neural networks is that they only need a previous value of the learning curve during operation and can use their internal short / long-term memory to remember values of the learning curve obtained from previous calculations. This means that the recurrent neural networks manage with little data during operation in order to be able to make a particularly reliable forecast. The advantage of the Bayesian recurrent neural network is that it can make a statement about the reliability of the prediction.

It is also proposed that the parameterization is processed by means of a plurality of fully connected neural networks and that the processed parameterizations are additionally provided as an input variable to one of the LSTMs. In addition, the processed parameterizations can be weighted depending on a randomly drawn variable, in particular from a Bernoulli distribution. This allows the parameterization of the learning process to be weighted, as a result of which the parameterization is fed to the LSTM according to its reliability. Another advantage of the fully meshed neural networks is their differentiability, which facilitates the optimization of new parameterizations. In addition, (fully meshed) neural networks can be trained with much more data, which improves the predictions.

It is also proposed that the majority of the LSTM of the Bayesian recurrent neural network and the learning curve are initialized before the start of the iterative learning process. When the learning curve is initialized, a first value is set equal to zero. The first value does not necessarily have to be zero; any values are possible as long as the same value is used when executing the method.

Depending on the parameterization and the value of the learning curve set to zero, the second machine learning system predicts the subsequent value of the learning curve, in particular without an iteration step of the learning process having to be carried out. It should be noted that this subsequent value of the learning curve corresponds to a value of the learning curve and it is expected that this would be present after the first iteration step of the learning method has been carried out. The advantage is that the iterative learning process does not have to be carried out, but a prediction of the learning curve can still be carried out. This allows a particularly efficient learning process which manages with a minimal number of iteration steps of the learning process. Initializing the Bayesian recurrent neural network can be understood to mean that all internal states of the recurrent neural network are set to a predetermined value, for example to the value zero.

It is also proposed that when the second machine learning system has predicted the subsequent value of the learning curve, the second machine learning system receives the predicted value of the learning curve as the previous value of the learning curve, and this step is repeated several times until a predeterminable plurality of values of the Learning curve was predicted. The advantage here is that a convergence of the learning process, in particular the learning curve, can be predicted.

It is further proposed that the iterative teaching method for teaching the first machine learning system is carried out, in particular continued, as a function of the prediction of the learning curve. It is also conceivable that, depending on the prediction of the learning curve, the learning method is parameterized with a different parameterization and the first machine learning system is then preferably learned with the other parameterization, or a learning curve is predicted for the other parameterization. If, for example, the learning curve does not reach a predefinable threshold value, then the parameterization can be discarded and a new parameterization selected. If the predicted learning curve exceeds the predefinable threshold value, then the iterative learning process is parameterized according to the parameterization and the first machine learning system is then learned with this parameterized learning process.

It is further proposed that the steps of the method are each carried out for a plurality of different parameterizations, and that the teaching method for teaching the first machine learning system is parameterized as a function of the predictions. This procedure allows an optimal parameterization of the iterative learning process to be explored.

It is also proposed that the second machine learning system depends on a A plurality of training data is learned and the training data comprise a plurality of training learning curves and parameterization assigned to each training learning curve. The training learning curves can be of different lengths.

It should be noted that the training data preferably originate from previous learning processes with similar learning processes and / or cost functions. Different length means that the training learning curves contain different numbers of values. It is conceivable that the second machine learning system is taught in parallel to the teaching of the first machine learning system. This has the advantage that the already generated learning curves of the first machine learning system can be used as training data for teaching the second machine learning system.

It is also proposed that the training data have data that were recorded when machine learning systems were trained for different tasks or applications.

As an alternative or in addition to the parameterization of the training method, metadata of the training data are preferably used. The advantage here is that this enables direct learning across different training data. Metadata can e.g. Be variables that characterize an argumentation of the training data (rotation / translation, etc.). Alternatively, instead of the metadata, a latent vector can be learned which is characteristic of the task which the first machine learning system is to solve after the learning. It is suggested that this vector can then be estimated based on a few observed data points of the learning curve on the new data set. For this purpose, i.a. Methods from the field of "Variational Inference" (e.g. Markov Chain Monte Carlo Method - MCMC for short - or Stochastic Variational Infence - SVI for short) can be used.

The invention of the first aspect can also be used for "bandit-based hyperparameter optimizers", e.g. BOHB, can be used to initialize this sensibly.

Since the invention of the first aspect is based on its advantages (such as the reliable, particularly inexpensive predictions of the learning curve and the outputting of probabilistic reliabilities of the predictions), the invention can also be used in "exploitation-exploration paradigm" scenarios, e.g. in reinforcement learning applications. Here, for example, the value of the learning curve can be a reward. This has the advantage that, in the case of rewards that are complex to observe or measure, they can easily be predicted using the second machine learning system.

In a second aspect of the invention, the first machine learning system is trained using the first aspect of the invention. The first machine learning system can then be used to control 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 of the first machine learning system can be a variable that was recorded by means of a sensor. A control variable can be determined as a function of an output variable of the machine learning system.

The control variable can be used to control an actuator of a technical system.

In a further aspect, a computer program is proposed. The computer program is set up to carry out one of the aforementioned methods. The computer program comprises instructions which cause a computer to carry out one of these named methods with all of its steps when the computer program runs on the computer. Furthermore, a machine-readable memory module is proposed, on which the computer program is stored. Furthermore, a device is proposed which is set up to carry out one of the methods.

Exemplary embodiments of the above-mentioned aspects are shown in the accompanying drawings and explained in more detail in the description below. Show:

Figure list

• 1 a schematic representation of a first machine learning system and a second machine learning system;
• 2 a schematic representation of an architecture of the second machine learning system;
• 3 a schematic representation of a flow chart of an embodiment of the method for predicting a learning curve;
• 4th a schematic representation of an embodiment of a device which can be used for teaching the first or second machine learning system.

1 shows schematically a first machine learning system ( 10 ), which by means of a device ( 11 ) is learned. The device ( 11 ) receives a training input variable, which by means of the first machine learning system ( 10 ) is processed into an output variable. The training input variables can, as exemplified in 1 depicted being pictures. However, other signals, such as audio signals, are also conceivable. After the output variable of the first machine learning system ( 10 ), a cost function (loss function) is used to determine a difference between the determined output variable and the training output variable depending on the determined output variable and a training output variable provided for the training input variable. The cost function then outputs a value y _t-1 . When the first machine learning system ( 10 ) is a deep neural network, then the cost function can be a square error, for example.

This value y _t-1 can then be used to parameterize the first machine learning system using an iterative learning process ( 10 ) so that the cost function is optimized, for example minimized.

The iterative learning process can be carried out several times in succession. The value of the cost function can be entered in a graph for each iteration step in order to map the course of the learning process. The reference number 13a shows an example of a diagram of a learning curve ( 13 ) of the first machine learning system ( 10 ). The individual iteration steps at time (t) are plotted on the x-axis and the values of the cost function are plotted on the y-axis. The diagram shows an example of how the values of the cost function behave along the iteration steps of the learning process.

Alternatively, the value y _{t-1 can} also be a performance of the first machine learning system ( 10 ) characterize on the training data, e.g. a classification accuracy of the images. It is then conceivable that a classification accuracy is plotted on the y-axis of the diagram and the learning curve ( 13 ) the convergence of the classification accuracy when teaching the first machine learning system ( 10 ) represents.

Since the iterative learning process can be parameterized, for example by a learning rate, gradient step size or the like, a convergence property of the learning curve ( 13 ) to be influenced. Since it is not possible to determine an optimal parameterization θ analytically, this is usually determined by manual trial and error. This is disadvantageous, since various parameterizations θ have to be tried out for this, most of which are discarded since they are not suitable for achieving a high level of performance in the first machine learning system. It is also questionable whether an approximately optimal parameterization θ can be found at all through manual trial and error if a parameterization space is multi-dimensional.

It is therefore proposed to use a second machine learning system ( 12 ), which receives the parameterization θ of the learning process and at least one value (y _t-1 ) of the learning curve as the input variable. The value (y _t-1 ) can have been determined for a previous iteration step (t-1) of the learning method and a performance of the first machine learning system ( 10 ) or be a value of the cost function. The second machine learning system ( 12 ) is learned in such a way that it has a subsequent value y _{t of} the learning curve ( 13 ) predicts which would have been determined after a subsequent iteration step (t) of the learning process carried out. This allows an initial assessment of whether, on the basis of the selected parameterization of the learning process, a useful learning of the first machine learning system ( 10 ) can be achieved.

The following is the learning curve ( 13 ) is viewed as a sequential time series of values of the cost function and / or the performance and a value (y _t ) of the learning curve is calculated based on previous observed values (y _tK-1 , ..., y _t-1 ) at previous iteration / Time steps (t - K - 1, ..., t - 1) predicted.

dabei entspricht Ψ einer Gauß-Verteilung und x_t = [θ,y_t-K-1, ...,y_t-1] und D beschreibt die Trainingsdaten, die insgesamt N unterschiedliche Trainingslernkurven mit unterschiedlicher Länge zusammen mit der jeweiligen zugeordneten Parametrisierung θ umfassen.The predicted value (y _t ) of the learning curve ( 13 ) can be determined as follows, assuming that the values (y _tK-1 , ..., y _t-1 ) were observed with _noise : $$p\left(y_t^{*}|x_t,\mathrm{D.}\right)=\Psi \left(\mu \left(y_t^{*}|x_t,\mathrm{D.}\right),{\sigma }^2\left(y_t^{*}|x_t,\mathrm{D.}\right)\right)$$

where Ψ corresponds to a Gaussian distribution and x _t = [θ, y _tK-1 , ..., y _t-1 ] and D describes the training _{data that comprise} a total of N different training _{learning curves of} different lengths together with the respective assigned parameterization θ .

wobei $\mu \left(y_t^{*}|x_t,D\right)={\displaystyle {\sum }_i{\mu }^i}$

und µⁱ entspricht den einzelnen Vorhersagen der Entscheidungsbäume und wobei ${\sigma }^2\left(y_t^{*}|x_t,D\right)=\frac{1}{N}{{\displaystyle {\sum }_i{\sigma }_i^2+\frac{1}{N}{\displaystyle {\sum }_i\left[{\mu }_i-\mu \left(y_t^{*}|x_t,D\right)\right]}}}^2,$

insbesondere unter Verwendung des Gesetztes der totalen Varianz (engl. law of total variance) berechnet wird.The second machine learning system ( 12 ) can be implemented in a first embodiment as a random forest with i-decision trees. The random forest determines the Gaussian distribution for every x _t $\Psi \left(\mu \left(y_t^{*}|x_t,\mathrm{D.}\right),{\sigma }^2\left(y_t^{*}|x_t,\mathrm{D.}\right)\right),$

in which $\mu \left(y_t^{*}|x_t,\mathrm{D.}\right)={\displaystyle {\sum }_i{\mu }^i}$

and µ ⁱ corresponds to the individual predictions of the decision trees and where ${\sigma }^2\left(y_t^{*}|x_t,\mathrm{D.}\right)=\frac{1}{N}{{\displaystyle {\sum }_i{\sigma }_i^2+\frac{1}{N}{\displaystyle {\sum }_i\left[{\mu }_i-\mu \left(y_t^{*}|x_t,\mathrm{D.}\right)\right]}}}^2,$

in particular using the law of total variance.

In the first embodiment, the random forest receives a total of K values (y _tK-1 , ..., y _t-1 ) of the learning curve ( 13 ) a previously unobserved learning curve. The random forest then tells the further course (y _t , ..., y _{t + X} ) of the learning curve ( 13 ) based on these total K values beforehand. This prediction of K values of the further course of the learning curve ( 13 ) will be referred to as a rollout in the following.

For a single rollout, the predicted value (y _{t + 1} ) is drawn from the determined Gaussian distribution Ψ and added to the determined values of the learning curve or replaces one of the determined values of the learning curve. This procedure can be repeated several times until a sequence with a predefinable length has been achieved.

und ${\sigma }^2\left(y_t^{*}|x_t,D\right)=\frac{1}{R}{\displaystyle {\sum }_r^R{\left(y_t^r-\mu \right)}^2},$

um eine zuverlässigere Vorhersage zu erhalten.In a further development of the first embodiment, several rollouts can be carried out for the total of K successively determined values (y _tK-1 ,..., Y _t-1 ) of the learning curve. All rollouts can then be averaged, for example: $\mu \left(y_t^{*}|x_t,\mathrm{D.}\right)=\frac{1}{\mathrm{R.}}{\displaystyle {\sum }_i^{\mathrm{R.}}{y}_t^r}$

and ${\sigma }^2\left(y_t^{*}|x_t,\mathrm{D.}\right)=\frac{1}{\mathrm{R.}}{\displaystyle {\sum }_r^{\mathrm{R.}}{\left(y_t^r-\mu \right)}^2},$

to get a more reliable prediction.

In a second embodiment of the second machine learning system ( 12th ) is this through a Bayesian recurrent neural network ( 20th ) given. A structure of the Bayesian recurrent neural network ( 20th ) is an example in 2 shown. The Bayesian recurrent neural network ( 20th ) receives a previously determined value y _{t-1 of} the learning curve ( 13 ) and processes this using LSTM layers connected in series ( 22nd , engl. LSTM layer). In the embodiment according to 2 includes the Bayesian recurrent neural network ( 20th ) two LSTM (r1, r2). In addition, the Bayesian recurrent neural network ( 20th ) at its exit with a fully meshed neural network ( 23 , engl. fully connected neural network), which has the value y _{t of} the learning curve ( 13 ) outputs.

In addition, the Bayesian recurrent neural network ( 20th ) at least one further fully meshed neural network ( 21st ), which processes the parameterization θ to h ₁ and also provides one of the LSTMs as an input variable. In 2 includes the Bayesian recurrent neural network ( 20th ) a plurality of further fully meshed neural networks ( 21st ).

In addition, the parameterization θ processed to h ₁ can be done by means of the further fully meshed neural network ( 21st ) are weighted depending on a quantity z, in 2 represented by: h ₁ * z ₁ . The quantity z can be randomly drawn from a Bernoulli distribution B (d). It should be noted that a parameterization d of the Bernoulli distribution B (d) when teaching the Bayesian recurrent neural network ( 20th ) can be optimized.

As already explained in the first embodiment, a rollout can be carried out here. Depending on the predicted value y _t , the Bayesian recurrent neural network ( 20th ) a subsequent value y _{t + 1} depending on the already predicted value y _t can be predicted, etc.

3 shows a schematic flow diagram of a method ( 30th ) to predict the learning curve ( 13 ).

The procedure ( 30th ) starts with step 3000 . In this step the second machine learning system ( 12th ) learned. For this purpose, the training data D is provided and then used to create the second machine learning system ( 12th ) to parameterize by teaching in such a way that the values y of the learning curve ( 13 ), especially from the training data.

After the second machine learning system ( 12th ) has been learned, step follows 3100 . In this step the second machine learning system is initialized. For example, if the second machine learning system is the Bayesian recurrent neural network ( 20th ), the internal states of the LSTM (r1, r2) can be initialized with the value zero, for example.

In the next step 3200 a parameterization θ of the learning process is selected and the first machine learning system ( 10 ) learned using this learning process. Depending on a value y _{t-1 of} the cost function, the second machine learning system ( 12th ) a value y _{t of} the cost function can be predicted (as shown in 2 ). Alternatively, the values y can represent a performance of the first machine learning system ( 10 ) characterize.

Alternatively, in step 3200 the first value y _{0 of} the cost function is initially set to the value 0 be set. It is conceivable that by then the first machine learning system ( 10 ) has not yet been learned.

In the next step 3300 the value y _t-1 or the first value y ₀ = 0 is determined by the second machine learning system ( 12th ) processed. The second machine learning system ( 12th ) then outputs the value y _t .

The following step is optional 3400 executed. The cost function from step is here dependent on the predicted value y _t 3300 another value y _{t + 1 of} the cost function predicted. This step 3400 can be executed several times in a row until a sequence y _t , ..., y _{t + X} of predicted values of the cost function has been determined.

In the next step 3500 becomes dependent on the predicted values after step 3400 or depending on the predicted value after step 3300 decided whether the parameterization θ will continue to be used to create the first machine learning system ( 10 ) or whether another parameterization θ * should be selected. For example, the other parameterization θ * can be drawn at random from a predefinable range of values. It is then conceivable that step again 3200 is carried out with the newly selected parameterization θ * in order to check whether the newly selected parameterization θ * is better suited.

After step 3500 is completed, step follows 3600 . In this step, the learning process is initialized with the parameterization θ and the first machine learning system ( 10 ) learned using this initialized learning process.

After step 3600 completed, can step 3700 consequences. In this step, the first learned machine learning system ( 10 ) for a technical system, such as an at least partially autonomous robot.

In an alternative embodiment of the method ( 30th ) the second machine learning system ( 12th ) are used in order to execute several rollouts depending on a plurality of different parameterizations θ provided and for each of the parameterizations with the initially selected value of the cost function equal to zero. Depending on the rollout that predicts the most suitable learning curve, the parameterization used for this is then selected and then used to train the first machine learning system ( 10 ) is used.

The at least partially autonomous robot can be an at least partially autonomous vehicle. In a further exemplary embodiment, the at least partially autonomous robot can be a service, assembly or stationary production robot, alternatively an autonomous flying object such as a drone.

The at least partially autonomous vehicle can comprise a detection unit. The detection unit can be, for example, a camera that detects the surroundings of the vehicle. The registration unit can use the first machine learning system ( 10 ) be connected. The first machine learning system ( 10 ) determines an output variable as a function of a provided input variable, for example provided by the acquisition unit, and as a function of a plurality of parameters of the first machine learning system. The output variable can be forwarded to a control unit.

The control unit controls an actuator as a function of the output variable of the machine learning system; it preferably controls the actuator in such a way that the vehicle executes a collision-free maneuver. In the first exemplary embodiment, the actuator can be a motor or a braking system of the vehicle. In a further exemplary embodiment, the partially autonomous robot can be a tool, a machine tool or a production robot. A material of a workpiece can be classified using the first machine learning system. The actuator can e.g. a motor that drives a grinding head.

4th shows a schematic representation of the device ( 11 ) to learn the first machine learning system ( 10 ), especially to carry out the steps for teaching the machine learning system ( 12th ) after the parameterization θ of the teach-in process has been initialized or adapted. The device ( 40 ) includes a training module ( 41 ), which contains the training data (images and assigned labels y _s ). Appropriately, in addition to images, sound sequences, text excerpts, radar, lidar or ultrasonic signals can be used, which are each labeled. Depending on the determined output variable y of the first machine learning system ( 10 ) and the labels y _s determine a differential module ( 42 ) a difference that is transferred to an adaptation module ( 43 ) is forwarded. The adaptation module ( 43 ) determines, depending on the difference, a change ϕ 'in the parameters ϕ of the first machine learning system ( 10 ), which are stored in a memory P, and are adapted depending on the change ϕ '.

The device can have a computing unit ( 44 ) and a storage element ( 45 ) include.

Claims (15)

Method (30) for predicting a learning curve (13) of an iterative learning method of a first machine learning system (10), the learning curve (13) comprising a plurality, in particular sequential time series, of values, the values of the learning curve (13) each having an iteration step of the iterative learning process are assigned and the values of the learning curve (13) each correspond to a value of a cost function of the respective iterative step of the iterative learning process or each a performance of the first machine learning system (10) on training data after executing the respective iterative step of the iterative Characterize the learning method, the iterative learning method being used to train the first machine learning system (10) and being parameterized by a hyperparameter θ, comprising the steps of: obtaining a value of the hyperparameter θ and at least one value (y _t-1 ) of the learning curve (13 ), the after at least one previously executed iteration step of the learning method was determined; and predictions, by means of a second machine learning system (12), depending on the value of the hyper parameter θ and depending on the at least one value (y _t-1 ) of the learning curve (13) of the, in particular immediately, subsequent value (y _t ) of the learning curve (13); and depending on the predicted value (y _t ) of the learning curve (13), deciding whether the value of the hyper parameter θ of the iterative learning process is maintained. Procedure according to Claim 1 , wherein the second machine learning system (12) is a random forest, the random forest depending on a plurality of values (y _t-1 , ..., y _tK-1 ) of the learning curve (13) the following value (y _t ) predicts the learning curve (13). Procedure according to Claim 1 , wherein the second machine learning system is a Bayesian recurrent neural network (20). Procedure according to Claim 3 , the Bayesian recurrent neural network (20) comprising a plurality of LSTMs (22) which are connected to one another in a predetermined sequence, the LSTMs (22) each determining their output variable depending on an input variable and their output variable determining their subsequent LSTM as an input variable, the input variable of the first LSTM of the plurality of LSTM being the at least one value of the learning curve (13). Procedure according to Claim 4 , the value of the hyper parameter θ being processed in each case by means of a plurality of fully connected neural networks (21, fully connected neural networks) and the processed value of the hyper parameter being additionally provided as an input variable to one of the LSTMs. Procedure according to Claim 5 , the processed value of the hyperparameters being weighted as a function of a randomly drawn variable, in particular from a Bernoulli distribution. Method according to one of the Claims 3 to 6th , the majority of the LSTM (22) of the Bayesian recurrent neural network (20) being initialized and the learning curve (13) initialized with a first predeterminable value, in particular equal to zero (y ₀ = 0), before the iterative learning method is used and wherein the second machine learning system (12) predicts its, in particular, subsequent value of the learning curve (13) depending on the value of the hyper parameter (θ) and depending on the first value (y ₀ = 0) of the learning curve (13). Method according to one of the preceding claims, wherein if the second machine learning system (12) has predicted the subsequent value of the learning curve (13), the second machine learning system (12) uses the predicted value of the learning curve (13) as the preceding value of the learning curve ( 13), and this step is repeated several times until a predeterminable plurality of values of the learning curve (13) has been predicted. Method according to one of the preceding claims, wherein the iterative teaching method for teaching the first machine learning system is carried out, in particular continued, as a function of the prediction of the learning curve (13). Procedure according to Claim 9 , wherein the steps of the method are each carried out for a plurality of different values of the hyperparameters, and wherein the training method for teaching the first machine learning system is parameterized depending on the predictions. Method according to one of the preceding claims, wherein the second machine learning system is trained as a function of a plurality of training data, the training data being a plurality of sequences (y ₁ , ..., y _N ) of the values of further learning curves and one associated with each sequence Include the value of the hyperparameters (θ ₁ , ..., θ _N ), with the further learning curves being recorded when training machine learning systems for different tasks. Procedure according to Claim 11 wherein the training data comprises a sequence of the values of the learning curve (13) of the learning of the first machine learning system (10). Computer program which comprises instructions which, when these instructions are carried out by means of a computer, cause the computer to carry out the method according to one of the preceding claims. Machine-readable storage medium (45) on which the computer program is based Claim 13 is stored. Apparatus (44) which is set up, the method according to any one of the preceding Claims 1 to 12th execute.

Images