Method for training machine learning algorithm in consideration of at least one inequality condition

Abstract

To provide a method for training a machine learning algorithm in consideration of at least one inequality condition, at least each one of inequality conditions representing one secondary condition.SOLUTION: A method (1) includes a step (2) of optimizing hyper parameters for a machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function adjusted on the basis of at least one inequality condition, and a step (3) of training the machine learning algorithm on the basis of the optimized hyper parameters.SELECTED DRAWING: Figure 1

Description

Technical Field

[0005]

[0001] The present invention relates to a method for training a machine learning algorithm considering at least one inequality condition, wherein each of the at least one inequality conditions represents a quadratic condition, and in particular, to a method for training a machine learning algorithm based on hyperparameters optimized considering at least one inequality condition.

Background Art

[0002] Machine learning algorithms generally rely on statistical methods to train a data processing system so that the data processing system can perform specific tasks that are not explicitly programmed for the data processing system. Here, the purpose of machine learning is to construct an algorithm that can learn from data and make predictions.

[0003] In this case, such machine learning algorithms are trained based on training data that characterizes each application case before or before application. Here, the weights within the machine learning algorithm are automatically adjusted so that the mapping of the relationship between features and predictions, or the relationship between input data and corresponding output data, can be made easier by the machine learning algorithm.

[0004] However, the structure of the machine learning algorithm defined through so-called hyperparameters, such as the architecture of the algorithm, is not affected by the training of the machine learning algorithm. Hyperparameters are parameters of a machine learning algorithm that do not need to be directly adjusted by training data or set before training, such as the number of layers of a neural network.

[0005] However, the quality or performance of a machine learning algorithm is also determined by its hyperparameters, so it is important to optimize these before the actual training of the machine learning algorithm.

[0006] Furthermore, when optimizing hyperparameters, it is important to consider peripheral conditions or quadratic conditions, such as settings related to available computer resources. In this case, such quadratic conditions often exist in the form of inequalities.

[0007] From U.S. Patent No. 1,1093833, a method for training a machine learning algorithm is known, wherein the training of the machine learning algorithm is based on tuned hyperparameter values. Here, if a selected hyperparameter configuration does not satisfy the linear constraint, it is determined whether a projection of the selected hyperparameter configuration is contained in a first cache that stores previously computed projections. If a projection is contained in the first cache, the projection is extracted from the first cache using the selected hyperparameter configuration, and the selected hyperparameter configuration is replaced by the extracted projection. If a projection is not contained in the first cache, a projection computed for the selected hyperparameter configuration is assigned to the session. The computed projection is received by the session for the selected hyperparameter configuration. The computed projection and the selected hyperparameter configuration are stored in the first cache, and the selected hyperparameter configuration is replaced by the computed projection. [Prior art documents] [Patent Documents]

[0008] [Patent Document 1] U.S. Patent No. 11093833 [Overview of the project] [Problems that the invention aims to solve]

[0009] Therefore, the problem underlying the present invention is to provide a method for training a machine learning algorithm that takes into account improved quadratic conditions that exist in the form of inequalities. [Means for solving the problem]

[0010] The problem is solved by the method described in the feature portion of claim 1 for training a machine learning algorithm while taking into account at least one inequality condition.

[0011] The problem can also be solved by a control device that trains a machine learning algorithm taking into account at least one inequality condition, as described in the feature portion of claim 6.

[0012] Disclosure of the invention According to one embodiment of the present invention, the above problem is solved by a method for training a machine learning algorithm considering at least one inequality condition, each of which expresses a quadratic condition, the method comprising the steps of optimizing hyperparameters for a machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function tuned on at least one inequality condition, and training the machine learning algorithm based on the optimized hyperparameters.

[0013] A tree-structured Parzen estimator is understood here as a method for processing category hyperparameters in a tree structure, or a method for generating a Parzen estimator in a search space with conditional hyperparameters. For example, a tree structure is required for selecting the number of layers in a neural network and the number of neurons in each layer. Furthermore, two distributions are defined for the hyperparameters or density, specifically the distribution when the output value of the objective function is less than a threshold, and the distribution when the output value of the objective function is greater than or equal to a threshold.

[0014] In this case, for example, the hyperparameters are divided into good and bad values. Furthermore, the objective function is a function that converts the hyperparameters to real values, which should be minimized within the framework of hyperparameter optimization. Subsequently, the two densities are modeled using a Parzen estimator or kernel density estimator, which represents the simple mean of kernels centered to existing data points. Here, a set of hyperparameters is output according to the expected maximum correction, or the ability to correct individual selections of hyperparameters is estimated.

[0015] Tree-structured Parzen estimators are characterized by their versatility and stable performance compared to other common methods for optimizing hyperparameters, such as evolutionary strategies, which are particularly distribution-based.

[0016] The acquisition function or selection function here further represents a function that defines the criteria for which the following set of hyperparameters is selected. Here, the criteria may be, for example, the expected modification.

[0017] In this case, the fact that the acquisition function is tuned based on at least one inequality has the advantage that the hyperparameters can be well optimized and the optimized hyperparameters are robust to at least one inequality, even if there are settings regarding the computing resources available for hyperparameter optimization or training of the machine learning algorithm, for example, if at least one inequality exists.

[0018] Thus, an improved method is provided for training machine learning algorithms to take into account quadratic conditions that exist in the form of inequalities.

[0019] In one embodiment, the method further includes the step of calculating an acquired function adjusted based on at least one inequality condition, wherein the step of calculating an acquired function adjusted based on at least one inequality condition includes the step of factorizing each of the at least one inequality condition.

[0020] Factorization, in this context, is understood as the decomposition of an object into multiple non-trivial factors. In particular, the mathematical definition of an inequality or quadratic condition can, in this case as well, be further processed by a tree-structured Parzen estimator and, in particular, the corresponding acquisition function, to decompose each into two distributions: one for the case where the output value of the corresponding objective function is less than a threshold, and another for the case where the output value of the objective function is greater than or equal to a threshold.

[0021] Thus, since a common combined distribution for the model and at least one inequality condition or quadratic condition can be selected as a basis for the optimization of hyperparameters, it can be ensured that the optimized hyperparameters remain robust with respect to at least one inequality condition. Considering only the distribution regarding at least one inequality condition further has the advantage that relatively few computer resources are required overall for the optimization of hyperparameters. Also, through the factorization or various distributions regarding at least one inequality condition, various observations regarding each of the at least one inequality condition can be incorporated into the optimization of hyperparameters.

[0022] Here, the step of calculating an acquisition function adjusted based on at least one inequality condition may further include the step of multiplying an acquisition function for the objective function and one acquisition function for each of the at least one inequality condition. Thereby, a common combined distribution for the model and at least one inequality condition can be calculated by a simple method and with relatively few computer resources.

[0023] At least one inequality condition may further be at least one setting regarding available computer resources. Therefore, the conditions of the data processing system in which the optimization of hyperparameters is implemented or executed are also themselves included together in the optimization of hyperparameters.

[0024] [[ID=1)) According to another embodiment of the present invention, there is also presented a method for classifying image data, wherein the image data is classified using a machine learning algorithm trained to classify the image data, and the machine learning algorithm is trained using the method of training a machine learning algorithm considering at least one inequality condition as described above.

[0025] A method for classifying image data based on a machine learning algorithm trained by a method of training a machine learning algorithm considering second-order conditions in the form of improved inequality conditions is presented. Here, the fact that the optimization of hyperparameters is based on an acquisition function adjusted based on at least one inequality condition means that when there is at least one inequality condition, for example, when there are settings regarding computer resources available for optimizing hyperparameters or for training a machine learning algorithm, the hyperparameters can be well optimized, and the optimized hyperparameters have the advantage of being robust to at least one inequality condition.

[0026] Here, in particular, the corresponding machine learning algorithm can be used to classify image data, especially digital image data, based on low-level features such as edges or pixel attributes. In this case, further, an image processing algorithm can be used to analyze the classification results aggregated into the corresponding low-level features.

[0027] According to another embodiment of the present invention, there is also provided a control device for training a machine learning algorithm considering at least one inequality condition, wherein each of the at least one inequality conditions represents a quadratic condition, and the control device is configured to optimize hyperparameters for the machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function adjusted based on at least one inequality condition, and a training unit configured to train the machine learning algorithm based on the optimized hyperparameters.

[0028] Thus, a control device is provided for training a machine learning algorithm while considering a quadratic condition that exists in the form of an improved inequality. The control device here is configured to optimize hyperparameters based on an acquisition function tuned on at least one inequality, which has the advantage that the hyperparameters can be well optimized even when at least one inequality exists, for example, when there are settings regarding the computing resources available for hyperparameter optimization or training the machine learning algorithm, and the optimized hyperparameters are robust to at least one inequality.

[0029] In one embodiment, the control device further includes a calculation unit configured to calculate an acquisition function adjusted based on at least one inequality condition, the calculation of the acquisition function adjusted based on at least one inequality condition includes factorization for each of the at least one inequality condition. Thus, as a whole, a common combined distribution for the model and at least one inequality condition or quadratic condition can be selected as the basis for hyperparameter optimization, thus ensuring that the optimized hyperparameters remain robust to at least one inequality condition. Considering only the distribution for at least one inequality condition further has the advantage that relatively few computer resources are required overall for hyperparameter optimization. Furthermore, various observations for each of the at least one inequality condition can be incorporated into the hyperparameter optimization by factorization or various distributions for at least one inequality condition.

[0030] Here, the calculation unit is further configured to calculate an acquisition function adjusted for at least one inequality condition by multiplying the acquisition function for the objective function by one acquisition function for each of the at least one inequality condition. This allows a common combined distribution for the model and at least one inequality condition to be calculated using a simple method and with relatively few computing resources.

[0031] At least one inequality condition may also be at least one setting regarding the available computing resources. Therefore, the conditions of the data processing system under which hyperparameter optimization is performed or carried out are themselves included in the hyperparameter optimization.

[0032] A further embodiment of the present invention also presents a control device for classifying image data, wherein the control device is configured to classify image data using a machine learning algorithm trained to classify image data, and the machine learning algorithm is trained by a control device that trains the machine learning algorithm taking into account at least one of the above-mentioned inequality conditions.

[0033] Thus, a control device is presented that classifies image data based on a machine learning algorithm trained by a control device that trains the machine learning algorithm considering a quadratic condition existing in the form of an improved inequality. In this case, the fact that the hyperparameter optimization is based on an acquisition function that is tuned based on settings for at least one inequality has the advantage that the hyperparameters can be well optimized even when there are settings for the computing resources available for hyperparameter optimization or training the machine learning algorithm, for example, when at least one inequality exists, and the optimized hyperparameters are robust to at least one inequality.

[0034] In particular, the corresponding machine learning algorithms can be used to classify image data, especially digital image data, based on low-level features, such as edge or pixel attributes. In this case, image processing algorithms can also be used to analyze the classification results aggregated to the corresponding low-level features.

[0035] In summary, the present invention provides a method for training a machine learning algorithm based on hyperparameters optimized to take into account at least one inequality condition.

[0036] The described configurations and developmental forms can be combined in any way.

[0037] Other possible configurations, variations, and embodiments of the present invention include combinations of features of the invention described above or below with respect to the examples that are not expressly mentioned.

[0038] The attached drawings are intended to provide a better understanding of embodiments of the present invention. These drawings illustrate embodiments and serve in connection with the description of the basic form and concept of the present invention.

[0039] Many other embodiments and the advantages described above become apparent in connection with the drawings. The elements shown in the drawings are not necessarily shown to scale relative to each other. [Brief explanation of the drawing]

[0040] [Figure 1] This is a flowchart of a method for training a machine learning algorithm considering at least one inequality condition according to an embodiment of the present invention. [Figure 2] This is a schematic block diagram of a control device for training a machine learning algorithm considering at least one inequality condition, according to an embodiment of the present invention.

[0041] In each drawing, identical or functionally equivalent elements, components, or parts are denoted by the same reference numeral unless otherwise specified. [Modes for carrying out the invention]

[0042] Figure 1 shows a flowchart of Method 1, an embodiment of the present invention, for training a machine learning algorithm while considering at least one inequality condition.

[0043] Here, each of the at least one inequality conditions expresses a quadratic condition.

[0044] Machine learning algorithms are based on two types of parameters, particularly hyperparameters and model parameters or weights. Model parameters can be learned, for example, during the training of a machine learning algorithm using labeled training data, while hyperparameters must be specified before training the machine learning algorithm.

[0045] Here, one way to select hyperparameters before training a machine learning algorithm is to manually search for the optimal hyperparameters, for example, by selecting the best possible hyperparameters based on empirical data and / or by manually testing various hyperparameters. Furthermore, the hyperparameters may be selected randomly or by random search, in which case the machine learning algorithm is subsequently trained based on the randomly selected hyperparameters.

[0046] For example, due to the increasing number and complexity of machine learning algorithms, methods for automatically selecting or automatically optimizing hyperparameters have also been developed. One example of such a method is Bayesian optimization. In Bayesian optimization, a probabilistic model is formed based on previous evaluation trials or previous selections of hyperparameters, and this probabilistic model maps the hyperparameters to the probabilities of evaluating the objective function. In this case, the hyperparameters are selected by optimizing the objective function.

[0047] In one evolution of Bayesian optimization, a tree-structured Parzen estimator is formed. Here, a tree-structured Parzen estimator is understood as a method for processing category hyperparameters in a tree structure, or a method for generating a Parzen estimator within a search space with conditional hyperparameters. For example, the tree structure is formed by selecting the number of layers in a neural network and the number of neurons in each layer. Furthermore, two distributions are defined for the hyperparameters or density, specifically, the distribution when the output value of the objective function is below a threshold and the distribution when the output value of the objective function is above a threshold. In this case, for example, the hyperparameters are divided into good and bad values. Here, the objective function is a function that converts the hyperparameters into real values, which should be minimized within the framework of hyperparameter optimization. Subsequently, the two densities are modeled using a Parzen estimator or kernel density estimator that represents the simple mean of kernels centered to existing data points. Here, a set of hyperparameters corresponding to the expected maximum modification is output, or the ability to modify individual selections of hyperparameters is estimated.

[0048] Here, the tree-structured Parzen estimator is characterized by its versatility and stable performance compared to other common methods for optimizing hyperparameters, such as evolutionary strategies, which are particularly distribution-based.

[0049] However, the tree-structured Parzen estimator has proven to be disadvantageous in that it has not been adapted to suit models that commonly appear in practical applications. Therefore, in models that commonly appear in practical applications, peripheral conditions or quadratic conditions, such as settings regarding available computer resources, are often considered. Such quadratic conditions often exist in the form of inequalities.

[0050] Here, Figure 1 illustrates Method 1, which includes step 2 optimizing hyperparameters for a machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function tuned on at least one inequality condition, and step 3 training the machine learning algorithm based on the optimized hyperparameters.

[0051] The fact that the acquisition function is tuned based on at least one inequality condition has the advantage that the hyperparameters can be well optimized and the optimized hyperparameters are robust to at least one inequality condition, even when there are settings regarding the computing resources available for hyperparameter optimization or training of the machine learning algorithm, for example, when at least one inequality condition exists.

[0052] Thus, overall, Method 1 is presented for training a machine learning algorithm to consider quadratic conditions that exist in the form of improved inequalities.

[0053] In particular, Method 1 is presented here, which represents an extension of the tree-structured Parzen estimator, in which the acquisition function is adjusted or extended based on at least one inequality condition.

[0054] Machine learning algorithms can be trained based on hyperparameters included in the configuration that maximizes the value of the acquisition function, which is calculated based on a tree-structured Parzen estimator.

[0055] Furthermore, the machine learning algorithm may be, for example, a neural network trained by deep learning. The hyperparameters to be optimized may be, for example, the number of layers in the neural network and the number of neurons per layer.

[0056] Furthermore, the method is generally applicable to black-box functions, that is, functions whose internal relationships are unknown, and only their input-output relationship, or a relationship purely defined by the mapping between input and output values, is known.

[0057] As shown in Figure 1, the illustrated method 1 further includes a step 4 of calculating an acquired function adjusted based on at least one inequality condition, the step of calculating an acquired function adjusted based on at least one inequality condition having a step of factorizing each of the at least one inequality condition.

[0058] In particular, here, for each of at least one inequality conditions, we can factorize one acquisition function or selection function for each corresponding inequality condition, that is, for example, we can form a distribution for good values ​​and a distribution for bad values, respectively.

[0059] According to the embodiment of Figure 1, step 4, which calculates the acquisition function adjusted based on at least one inequality condition, further comprises the step of multiplying the acquisition function for the objective function by one acquisition function for each of the at least one inequality condition, where the product of the acquisition function for the objective function and the acquisition functions for each individual inequality condition forms the acquisition function adjusted based on at least one inequality condition.

[0060] Therefore, the acquisition function, which is tuned based on at least one inequality condition, can be tuned to match the setup or configuration of individual inequalities or individual quadratic conditions. Furthermore, if there are no quadratic conditions expressed by the inequalities, hyperparameters are selected that have the same performance as a standard tree-structured Parzen estimator.

[0061] The "at least one inequality condition" refers to at least one setting relating to available computer resources, such as processor capacity, memory capacity, or latency.

[0062] A machine learning algorithm trained on appropriately selected or optimized hyperparameters can then be used, for example, for the classification of image data. In this case, the machine learning algorithm can be further trained on labeled comparison image data.

[0063] Furthermore, a reasonably trained machine learning algorithm can also be trained to control an autonomous vehicle based on a LiDAR model and / or radar model, even on reasonably labeled comparative image data, where the autonomous vehicle has resources that are typically limited, such as resources to optimize engine control or ABS control systems, or resources to optimize process parameters during the manufacturing of components, for example, during resistance welding, injection molding, or heat treatment of metals.

[0064] Figure 2 shows a schematic block diagram of a control device 10 that trains a machine learning algorithm considering at least one inequality condition, according to an embodiment of the present invention.

[0065] Here, each of the at least one inequality conditions also expresses a quadratic condition in this case.

[0066] As shown in Figure 2, the control unit 10 includes an optimization unit 11 configured to optimize hyperparameters for a machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function tuned on at least one inequality condition, and a training unit 12 configured to train the machine learning algorithm based on the optimized hyperparameters.

[0067] The optimization unit and the training unit can be implemented, for example, based on code stored in memory and executable by the processor.

[0068] Here, according to the embodiment of Figure 2, the control device 10 further has a calculation unit 13 configured to calculate an acquisition function adjusted based on at least one inequality condition, where the calculation of the acquisition function adjusted based on at least one inequality condition includes factorization of each of the at least one inequality condition.

[0069] In this case, the calculation unit can also be implemented, for example, based on code stored in memory and executable by the processor.

[0070] According to the embodiment shown in Figure 2, the calculation unit is further configured to calculate an acquired function adjusted for at least one inequality condition by multiplying the acquired function for the objective function by an acquired function for each of the at least one inequality condition.

[0071] At least one inequality condition is, in this case as well, at least one setting regarding available computer resources.

[0072] The control device 10 is configured, in particular, to perform a method for training a machine learning algorithm, taking into account at least one inequality condition as described above. Furthermore, the code for implementing the optimization unit, the code for implementing the training unit, and the code for implementing the calculation unit can also be integrated into a computer program product.

Claims

1. 1. A method for training a machine learning algorithm by taking into account at least one inequality condition, wherein each of the at least one inequality condition represents a respective second-order condition, the method comprising: (2) optimizing hyperparameters for the machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function adjusted based on the at least one inequality condition; (3) training the machine learning algorithm based on the optimized hyperparameters; A method comprising:

2. The method (1) further comprises the step (4) of calculating an adjusted acquisition function based on the at least one inequality condition; the step (4) of calculating the adjusted acquisition function based on the at least one inequality condition comprises factoring each of the at least one inequality condition; The method of claim 1.

3. the step (4) of calculating an adjusted acquisition function based on the at least one inequality condition includes a step of multiplying an acquisition function for the objective function by a respective acquisition function for each of the at least one inequality condition; The method of claim 2.

4. the at least one inequality condition is at least one setting regarding available computer resources; The method of claim 1.

5. 1. A method for classifying image data, comprising:

5. A method in which image data is classified using a machine learning algorithm trained to classify image data, the machine learning algorithm having been trained using the method for training a machine learning algorithm taking into account at least one inequality condition according to claim 1 .

6. 1. A control device for training a machine learning algorithm by taking into account at least one inequality condition, comprising: each of the at least one inequality condition expresses a respective second-order condition; The control device (10) an optimization unit (11) configured to optimize hyperparameters for the machine learning algorithm by applying a tree-structured Parzen estimator based on an acquisition function adjusted based on the at least one inequality condition; a training unit (12) configured to train the machine learning algorithm based on the optimized hyperparameters; A control device having:

7. the control device further comprises a calculation unit (13) configured to calculate an adjusted acquisition function based on the at least one inequality condition, wherein the calculation of the adjusted acquisition function based on the at least one inequality condition comprises a factorization of each of the at least one inequality condition. The control device according to claim 6.

8. the calculation unit (13) is further configured to calculate an adjusted acquisition function based on the at least one inequality condition by multiplying an acquisition function for the objective function by a respective acquisition function for each of the at least one inequality condition. The control device according to claim 7.

9. the at least one inequality condition is at least one setting regarding available computer resources; The control device according to claim 6.

10. A control device for classifying image data, comprising: the controller is configured to classify the image data using a machine learning algorithm trained to classify the image data; The machine learning algorithm is trained by a control device that trains the machine learning algorithm by taking into account at least one inequality condition according to any one of claims 6 to 9. Control device.

Images