WO2020187394A1 - Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé - Google Patents

Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé Download PDF

Info

Publication number
WO2020187394A1
WO2020187394A1 PCT/EP2019/056621 EP2019056621W WO2020187394A1 WO 2020187394 A1 WO2020187394 A1 WO 2020187394A1 EP 2019056621 W EP2019056621 W EP 2019056621W WO 2020187394 A1 WO2020187394 A1 WO 2020187394A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
input
type
inputs
code
Prior art date
Application number
PCT/EP2019/056621
Other languages
German (de)
English (en)
Inventor
Bodo Rosenhahn
Bastian WANDT
Original Assignee
Gottfried Wilhelm Leibniz Universität Hannover
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Gottfried Wilhelm Leibniz Universität Hannover filed Critical Gottfried Wilhelm Leibniz Universität Hannover
Priority to PCT/EP2019/056621 priority Critical patent/WO2020187394A1/fr
Publication of WO2020187394A1 publication Critical patent/WO2020187394A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Definitions

  • the invention relates to a method for training an autoencoder which has an encoder and a decoder.
  • the autoencoder can be trained in such a way that the encoder generates a code from an input representing the input and the decoder generates an output from the code which is a reconstruction of the input.
  • the invention also relates to an autoencoder trained using such a method and an encoder and a decoder of such an autoencoder.
  • the invention also relates to a method for classifying data.
  • the invention also relates to a computer program with program code means that is set up to carry out one of the methods mentioned.
  • An autoencoder is an artificial neural network that has an input layer (input layer), an output layer (output layer) and at least one hidden layer, which connects the input layer with the output layer.
  • a hidden layer of the autoencoder is designed as a code layer.
  • Each of the layers has a number of artificial neurons (hereinafter referred to as neurons for short).
  • the number of neurons in the code layer of an autoencoder can in particular be smaller than the number of neurons in the input layer and / or the number of neurons in the output layer.
  • the input layer and the output layer can in particular have the same number of neurons.
  • the autoencoder has an encoder and a decoder.
  • the encoder comprises the input layer, the code layer and all hidden layers in between and the connections between these layers.
  • the decoder comprises the code layer, the output layer and all the hidden layers arranged in between and the connections between these layers.
  • An input can be supplied to the autoencoder via the input layer and the autoencoder can output an output via the output layer.
  • the code generated by the encoder is formed by the outputs of the neurons of the code layer, i. H. by activating the neurons of the code layer.
  • the codes generated by the encoder are also known as latent variables. The space they create is also known as latent space.
  • a code can be understood as a position in the latent space of the auto coder. The generation of a code from an input is accordingly also referred to as projection into latent space.
  • the autoencoder can be trained in such a way that the encoder generates a code representing the input from an input and the decoder generates an output from the code which is a reconstruction of the input.
  • the autoencoder is usually trained using a training data set and an error function, which is also referred to as the loss function (loess function).
  • Autoencoders are generally known from the prior art and are usually used to learn efficient coding by means of unsupervised learning, i.e. a compressed representation of the data fed as inputs to the autoencoder. This is usually achieved by making the code layer a smaller dimension compared to the input layer and the output layer, i. has a smaller number of neurons. The artificial neural network is therefore forced to learn to generate a code that is an efficiently compressed representation of the input. Autoencoder can, for example, be used to denoise
  • a method for processing image data with the aid of an auto-encoder is known from US 2017/0076224 A1. It is intended that the autoencoder training is carried out using a training data set with positive training data and a training data set with negative training data in such a way that the autoencoder reconstructs the positive training data but does not reconstruct the negative training data. In this way, the noise in the image data and / or other undesirable components in the image data can be prevented from being reproduced by the auto-encoder.
  • the object of the present invention is therefore to improve the auto-encoders known from the prior art in such a way that the disadvantages explained above are avoided, and on this basis to provide improved options for using auto-encoders.
  • an error function of the first type which comprises a first error term and a second error term, the first error term describing a reconstruction error in the form of a deviation between a training input and the output generated therefrom, and the second error term describing a deviation between an actual value generated from the training input Code and the target code assigned to the training input,
  • any type of autoencoder with any structure can be used as the autoencoder that is trained with the method according to the invention.
  • the encoder and / or the decoder of the autoencoder can be used as a convolutional neural network with any number of layers and any number of filters per layer and any activation functions be trained.
  • the encoder and / or the decoder of the car encoder can also be used, for example, as a fully connected neural network (fully
  • the dimension of latent space i.e. the number of neurons in the code layer can in principle be chosen as desired.
  • the inputs used to carry out the method, in particular the training inputs can in principle be data of any type.
  • the inputs, in particular the training inputs can in particular be image data.
  • the inputs, in particular the training inputs can, however, e.g. also audio data and / or video data and / or text data and / or other data be th.
  • Training inputs of the first type in the context of the present application are training data to which a known expression of at least one attribute is assigned.
  • the attribute can basically be of any type.
  • An example of such an attribute can be, for example, the seasonal determination of items of clothing that are shown in the image data of the first type that form the training inputs of the first type.
  • Such an attribute could be winter clothing, summer clothing and all-season clothing, for example.
  • a further example of an attribute can be the gender of a person who is shown in the image data forming the training inputs of the first type.
  • Such an attribute could, for example, be masculine or feminine.
  • Another example of an attribute can be the parity of a number shown in the image data, the attribute in this case being able to have the values even or odd.
  • the structure of the inputs ie the characteristics of the at least one attribute
  • the at least one attribute and its expression, which is assigned to each training input of the first type can in principle however be determined arbitrarily, ie the assignment of the known expression of the at least one attribute to the training inputs of the first type can take place according to arbitrarily determined assignment rules.
  • At least some of the training inputs of the training data set with which the autoencoder is trained are each provided with a label, i.e. are labeled training inputs.
  • the label indicates the known expression of the attribute.
  • a target code assigned to the respective training input is determined.
  • the target code is determined as a function of the characteristics of the attribute that is assigned to the respective training input.
  • the target code is a code of the type explained above.
  • the target code can accordingly be understood as the target position in the latent space of the autoencoder.
  • the autoencoder is trained using at least one error function of the first type.
  • the error function of the first type is an error function that includes a first error term and a second error term.
  • the first error term describes a reconstruction error in a manner known per se, i.e. a discrepancy between a training input that is supplied to the Autoenco via the input layer and the output that the Autoencoder generates from this input and outputs via the output layer. This output generated from the input is therefore the output that the decoder generates from the code representing the input.
  • the second error term describes a discrepancy between an actual code generated by the encoder from the training input that is fed to the autoencoder via the input layer and the target code assigned to the same training input.
  • the actual code is also a code that is to be before the explained type and can therefore be understood as the actual position in the latent space of the car encoder.
  • the actual code is a code that the encoder actually generates from an input.
  • the error function of the first type can, for example, comprise an addition of the first error term and the second error term or consist of such an addition.
  • the first error term and the second error term can accordingly be linked in the error function, for example by addition.
  • the first error term and the second error term can, however, also be linked in other ways in the error function of the first type, e.g. by a multiplication.
  • the error function of the first type can contain other terms in addition to the first error term and the second error term.
  • the training data record can also include training inputs for which the expression of an attribute is not known.
  • the code in particular the target code and / or the actual code, can in particular be a deterministic code, i.e. be a deterministic position in latent space.
  • the invention offers the advantage that a structure that is present in the input data and that results from the various characteristics of one or more attributes of the inputs is retained when the inputs are projected into latent space, ie when the codes are generated from the inputs what remains is that it is reflected in the structure of the codes in latent space.
  • This structuring of the codes in latent space is achieved according to the invention in that the error function of the first type has the second error term, which describes the deviation between the actual code and the target code determined depending on the characteristics of the attribute.
  • the second error term describes a structuring error in the codes in latent space.
  • a large discrepancy between the actual code and the target code, that is to say a large structuring error has the consequence that the error function of the first type indicates a large error.
  • This error can occur in the context of learning the artificial neural network, ie when training the autoencoder, for example by means of a backpro pagation method (error feedback) to adapt the weights of the artificial neural network so that the error indicated by the error function of the first type, which is both the reconstruction error and the second error term Structuring errors described includes, minimized or at least reduced.
  • the actual codes generated by the encoder approximate the target codes, which are dependent on the characteristics of the attribute, through the training of the autoencoder.
  • the codes actually generated from the inputs have a structure in the form of a distribution of the codes in latent space, which reflects the structure of the inputs, ie the characteristics of the attributes assigned to them. In this way, it is possible to separate the codes generated from the inputs in latent space according to the respective expression of the attribute.
  • the first error term which describes the reconstruction error, is achieved in a manner known per se that when the artificial neural network is learned, i.e. when training the autoencoder, the reconstruction error is minimized or at least reduced.
  • an autoencoder trained according to the invention is able to generate a structure of the codes in latent space which excellently reflects the characteristics of the attributes of the associated inputs .
  • the structure created in this way in the latent space can be used in a variety of ways.
  • the generated structure of the codes ie their distribution in latent space, can be used for a simple and reliable classification of the data with regard to the expression of the attribute.
  • the structure of the codes can also be used for a synthetic generation of data, in which outputs with a desired expression of the attribute can be generated by means of arbitrarily generated codes.
  • the training data set includes, in addition to the plurality of training inputs of the first type, a plurality of training inputs of the second type to which no known expression of the attribute is assigned, and the auto-encoder is trained using the error function of the first Type and an error function of the second type describing the reconstruction error.
  • the error function of the first type is applied to the training inputs of the first type and the error function of the second type is applied to the training inputs of the second type.
  • the training data set with which the autoencoder is trained includes, in addition to the training inputs of the first type, to each of which a known expression of the at least one attribute is assigned, a large number of training inputs of the second type, to which no known expression of the at least one attribute is assigned . Provision is therefore made for the autoencoder to be trained using a training data set which includes both labeled training inputs (training inputs of the first type) and unlabeled training inputs (training inputs of the second type).
  • an error function of the second type is used.
  • This error function of the second type describes the reconstruction error in the sense of a discrepancy between a training input and the output generated from it.
  • the error function of the second type can, however, be designed in such a way that it does not describe the structuring error in the sense of a deviation between an actual code generated from the training input and an actual code assigned to the same training input that is dependent on the characteristics of the attribute Target code.
  • the error function of the second type can, for example, include the first error term explained above, but not include the second error term explained above.
  • the first The error term can, for example, match the error function of the first type and the error function of the second type.
  • Such a further development of the invention offers the advantage that available unlabeled training inputs can be used to minimize or at least reduce the reconstruction error when training the auto-encoder.
  • the testing of the inventive method for training an autoencoder and the appropriately trained autoencoder has shown that very good results can be achieved if the proportion of training inputs of the first type in the total number of training inputs in the training data set is very low.
  • the training method according to the invention when processing a test data set with the aid of the autoencoder trained according to the invention, it is possible to achieve both good structuring of the codes in latent space and low reconstruction errors when the autoencoder has been trained with a training data set that contains le only contains a small proportion of labeled training inputs (training inputs of the first type).
  • the ratio n1 / (n1 + n2) is not greater than 0.2.
  • n1 denotes the number of training inputs of the first type, ie the number of labeled training inputs that are included in the training data set
  • n2 denotes the number of training inputs of the second type, ie the number of unlabeled training inputs that are included in the training data set.
  • This ratio can advantageously not be greater than 0.1 in particular.
  • This ratio can in particular advantageously not be greater than 0.05.
  • This ratio can in particular advantageously not be greater than 0.02. In particular, this ratio can advantageously not be greater than 0.01.
  • n1 + n2 can be interpreted as the total number of training inputs in the training data set.
  • This development of the invention thus provides that a training data set is used to train the auto encoder, which only has a small proportion of labeled training inputs, ie a small proportion of training inputs of the first type.
  • the label in the training input ie the assignment of the characteristics of the attribute to the training input of the first type, is complex, since it usually has to be done manually.
  • the proportion of labeled training inputs is kept low, offers the advantage according to the invention that the training data set can be provided with little effort and therefore at low cost.
  • the expression of the attribute describes an association with a class.
  • the attribute can be a classification criterion.
  • the attribute is the seasonal determination of the respective item of clothing and its characteristics describe whether it belongs to one of the classes of summer clothing, winter clothing or year-round clothing.
  • an autoencoder trained according to the invention structures the codes in latent space in such a way that the structure of the codes reflects the various classes.
  • An autoencoder trained according to the invention - or just its encoder - can in this way be used particularly effectively for applications in the field of data classification.
  • an autoencoder trained according to the invention - or just its decoder - can be used in this way to generate synthetic outputs of a certain class and / or to generate synthetic outputs that are a result of an interpolation between the classes.
  • the determination of the target code, which is assigned to the respective training input takes place in that a target code is determined for each characteristic of the attribute and that target code is assigned to each training input of the first type , which was determined for the characteristic of the attribute assigned to it.
  • each expression of the attribute is assigned a target code so that the target code for each training input of the first type can be determined solely on the basis of the characteristic of the attribute assigned to it.
  • Such a development of the invention offers the advantage that the target codes for the training inputs of the first type can be determined in a particularly simple manner, since for this purpose only the characteristic of the attribute assigned to the respective training input has to be checked.
  • the reference codes are determined manually.
  • the target codes can be determined manually, in particular, in that each training input of the first type is assigned an arbitrarily determined target code depending on the characteristic of the attribute assigned to the training input. For example, an arbitrary target code, ie an arbitrary position in latent space, can be determined for each occurrence of the attribute in the total amount of training inputs of the first type. On this basis, each training input of the first type can then be assigned the target code that was arbitrarily determined for the characteristic of the attribute assigned to it.
  • the target codes can in particular be determined in such a way that the deviations between the different target codes, for example measured by their distances in latent space, correspond to the similarities and / or dissimilarities between the attributes of the attribute.
  • similar manifestations of the target code attribute can be determined, which differ from one another only comparatively slightly, ie in the latent space have a small distance from one another, while for one of the very dissimilar characteristics of the attribute target codes can be determined which differ greatly from one another, ie are a large distance from one another in the latent space.
  • a structuring of the codes in the latent space is advantageously achieved, which reflects the similarities or dissimilarities between the characteristics of the attribute.
  • a distance between the two expressions is determined for one or more combinations of two expressions of the attribute and the target codes are determined as a function of the distances between the expressions.
  • the target codes are determined as a function of the distances between the training inputs. For this purpose, a distance between the two values is also determined for one or more combinations of two values of the attribute. In addition, one or more combinations of training inputs of the first type, each consisting of a first training input with a first expression of the attribute and a second training input with a second expression of the attribute, will be the distance between the first expression as the distance between the training inputs and assigned to the second expression.
  • a distance between the two values can advantageously be determined for all combinations of two values of the attribute which the training inputs of the first type of the training data record have. It is also conceivable that a distance between the two values is determined for every possible combination of two values of the attribute.
  • the determination of the distances between the characteristics can in principle be done in an arbitrary manner. The distances between the characteristics can be determined manually, for example. In the simplest case, it is possible, for example, that the same distances are determined for all combinations of different characteristics.
  • the value dü 0, for example, can be selected for the distances between the same values of the attribute, with i in turn denoting the value of the attribute.
  • a low value can be determined for a distance between two characteristics that are similar to one another, and a high value can be determined for a distance between two characteristics that differ greatly from one another.
  • the distances between the expressions can in particular be determined in such a way that the distances between the expressions form a distance matrix, and the target codes can be determined as a function of this distance matrix.
  • the distances between training inputs can also be determined in such a way that the distances between the training inputs form a distance matrix, and the target codes can be determined as a function of this distance matrix.
  • Target codes from the distances between the values and / or from the distances between the training inputs can be determined using a method for dimension reduction.
  • the target dimension of the dimension reduction is the dimension of the latent space of the autoencoder.
  • a method for dimensional reduction can be applied to a distance matrix of the type discussed above; H. to a distance matrix formed from the distances between the characteristics and / or to a distance matrix formed from the distances between the training inputs.
  • the target codes are determined from the distances between the values and / or from the distances between the training inputs using one or more of the following methods or all of the following methods:
  • the methods mentioned can also be easily implemented in software and / or hardware, so that the determination of the reference codes can be automated without any problems.
  • Such developments of the invention therefore offer particular advantages when the number of different Ausgar conditions of the at least one attribute is large and therefore manual determination of the target codes is difficult or not easily possible.
  • Classifying each input to be classified by assigning a class to the respective input depending on the code representing the respective input.
  • Classifying each input to be classified by assigning a class to the respective input depending on the code representing the respective input.
  • the codes generated by an encoder of an autoencoder according to the invention are - as already explained above - arranged in the latent space in such a way that they are separated according to their respective characteristics of the attribute, ie the codes have a structure in the latent space that follows the characteristics of the attributes.
  • This property of the autoencoder trained according to the invention can be used to classify the data on the basis of the structure of the codes in latent space.
  • the class assigned to each input to be classified reflects the characteristic of the attribute assigned to it.
  • the input and the class assigned to it form a classification result.
  • Such a method according to the invention for classifying data offers the advantage that it offers a significant increase in classification reliability compared to comparable methods known from the prior art.
  • This is achieved by making use of the structuring of the codes in latent space for the classification of the data. Since the inventive training of the autoencoder ensures that the codes that are generated from inputs with different characteristics of the attribute are clearly separated from one another in latent space, the codes can be used to easily and reliably distinguish between different characteristics of the attribute and based on this A classification can be carried out on the basis of this.
  • the class is assigned using a machine classification method.
  • a machine classification method in the sense of the present application is understood to mean an automatic classification method.
  • Such a development of the invention offers the advantage that a machine classification method can be implemented as an implementation in software and / or hardware, so that the class can be assigned automatically and automatically. In this way, even a very large number of inputs to be classified can be classified quickly and reliably.
  • the class is assigned by a support vector machine (SVM).
  • SVM support vector machine
  • the support vector machine can in particular be a multi-class support vector machine be.
  • the class can be assigned using a k-Nearest-Neighbor method.
  • the class can be assigned using a random forest method.
  • the class can also be assigned using an artificial neural network set up for classification.
  • the reliability indicator shows the reliability of the classification decision.
  • a reliability indicator can be determined for a classified input, for example, as a function of the distance between the code representing the classified input and a class boundary in the has latent space.
  • the distance between the code representing the classified input and the class boundary can be used as a reliability indicator.
  • the reliability of the classification decision indicated by the reliability indicator is greater, the greater the distance from the class boundary.
  • Such a class boundary in latent space can be a hyperplane, for example.
  • Such a hyper level is determined as a class boundary, for example by a support vector machine.
  • the reliability indicator is determined on the basis of a distance between the code representing the classified input and at least one other code that represents another classified input of the same or a different class.
  • a distance to at least one other code can be, for example, a distance to a k-nearest neighbor if the class is assigned using a k-nearest neighbor method.
  • the reliability indicator can be determined, for example, as the distance to an adjacent code of the same class and / or to an adjacent code of a different class.
  • a short distance to an adjacent code of the same class can indicate a high level of reliability and a high distance to an adjacent code of the same class can indicate a low level of reliability.
  • a large distance to an adjacent code of a different class can indicate a high level of reliability and a small distance to an adjacent code of a different class can indicate a low level of reliability.
  • scoring values are determined for all classes that come into consideration for the assignment to an input to be classified.
  • the input to be classified is usually assigned the class for which the best scoring value was determined.
  • the reliability indicator is determined as a function of a scoring value that for the class assigned to the classified input was determined. Usually this is the best scoring value of all the relevant classes.
  • the reliability indicator for a classified input can therefore also be determined, for example, as a function of the best and the second-best scoring value determined for the classified input.
  • the reliability indicator can be determined as the difference between the best scoring value and the second-best scoring value, a large difference indicating a high reliability of the classification decision and a small difference indicating a low reliability of the classification decision.
  • one or more critical inputs can be selected from the classified inputs, for example, the critical inputs being selected as a function of the reliability indicators determined for the inputs. It is conceivable, for example, that a classified input is selected as a critical input when the reliability indicator exceeds or falls below a predetermined threshold value. It is also conceivable, for example, to sort the classified inputs according to the reliability indicator determined and to select those classified inputs as critical inputs for which the reliability indicator shows the lowest reliability of the classification decision. For this purpose, for example, a predefined number of critical inputs to be selected can be predefined and / or a predefined proportion of critical inputs to be selected in relation to the total number of classified inputs can be predefined. The selected critical inputs can then be used to improve classification reliability.
  • the value of the attribute is determined manually and assigned to the critical input.
  • at least one critical input can be assigned a class manually.
  • one or more critical inputs in particular all critical inputs, be labeled manually afterwards and / or that these critical inputs are manually assigned a class.
  • Such a development of the invention offers the advantage that, with the aid of the critical inputs post-processed in this way, the classification reliability can be significantly improved.
  • Another advantage of such a further development of the method according to the invention is that it is possible in this way to select specifically those inputs for subsequent manual labeling or manual classification for which the machine classification can only make a classification decision with low reliability .
  • Manual post-processing of the entries which is usually time-consuming and costly, can be reduced to a necessary minimum in this way, so that considerable savings can be achieved in this regard.
  • the critical entries must be labeled, ie each critical entry must be assigned a known expression of at least one attribute.
  • the expression of the attribute for each critical input that is to be used as training input of the first type manually determined in the manner explained above and assigned to the critical input.
  • Such a further development of the invention offers the advantage that by retraining the autoencoder in this way, the reliability with which the classification decisions can be made can be improved in a targeted and significant manner. This is achieved in that the retraining of the autoencoder takes place precisely with those inputs as training inputs of the first type for which the autoencoder could previously only make a classification decision with low reliability.
  • a code is determined and fed to the decoder of a car encoder trained according to the invention, so that the decoder generates an output from the code.
  • gen can.
  • the code can basically be determined arbitrarily.
  • the code can be a target code of the type explained above, e.g. B. a target code intended for an expression of the attribute.
  • the decoder of the autoencoder trained according to the invention it is possible to use the decoder of the autoencoder trained according to the invention to generate a synthetically generated output which corresponds to the characteristic of the attribute given by the reference code.
  • image data can be generated synthetically which, depending on the given code, show either a male or a female human body.
  • the code is determined by adding a transformation vector to a reference code.
  • the transformation vector is formed by scaling a difference vector between two different reference codes that were determined for different characteristics of the attribute when training the autoencoder.
  • the reference code can in particular be a target code.
  • the code can be determined, in particular, by interpolating between two different reference codes that were determined for different versions of the attribute when the auto-encoder was trained. Such an interpolation can take place, for example, in that one of the two different reference codes is selected as the reference code of the type explained above.
  • the scaling can for example be done by a multiplication with a scaling factor.
  • the scaling can in particular be a compression of the difference vector.
  • the scaling can also be a stretching of the difference vector.
  • the scaling factor can also have the value one, so that the transformation vector corresponds to the difference vector.
  • Such a development of the method according to the invention for generating a ner output offers the advantage that outputs can be generated synthetically, which with regard to the at least one attribute do not follow a predetermined expression of the attribute, but z. B. correspond to a mixture of different characteristics.
  • a mixture of different characteristics is achieved by adding the difference vector to the reference code.
  • Such a mixture between two forms can be achieved in particular by interpolation between two different reference codes.
  • a means is available to synthetically generate outputs with an arbitrary expression of the at least one attribute.
  • the object mentioned at the beginning is also achieved by a decoder of an autoencoder that has been trained with a method for training an autoencoder of the type explained above.
  • the above-mentioned object is also achieved by a computer program with program code means which is set up to carry out a method or several methods or all methods for training an autoencoder and / or for classifying data and / or for generating an output of the type explained above, if the computer program is executed on a computer.
  • Fig. 1 - a schematic representation of the structure of an autoencoder that can be trained with the method according to the invention
  • FIG. 3 shows part of an exemplary training data set with training inputs of the first type and of the second type
  • Figure 4a is a 2D projection of codes in latent space generated by an autoencoder trained in a conventional manner
  • FIG. 1 shows a schematic representation of the structure of an auto-encoder 1, which can be trained with a method according to the invention for training an auto-encoder.
  • the auto-encoder 1 has an input layer 11, an output layer 14 and a total of seven hidden layers 12a, 12b, 12c, 12d, 12e, 12f and 13 which connect the input layer 11 to the output layer 14.
  • One of these hidden layers is code layer 13.
  • An input can be fed to the autoencoder via the input layer 11 and the autoencoder can output an output via the output layer 14. ben.
  • the input layer and the output layer each have 28 ⁇ 28 neurons, that is, 784 neurons.
  • 28 x 28 pixel images can be fed to the autoencoder as inputs via the input layer and the autoencoder can output 28 x 28 pixel images via the output layer as outputs.
  • the auto encoder 1 has an encoder 3 and a decoder 5.
  • the encoder 3 comprises the input layer 11, the hidden layers 12a, 12b and 12c and the code layer 13 and the connections (not shown in FIG. 1) between these layers.
  • the decoder 5 comprises the code layer 13, the hidden layers 12d, 12e, 12f and the output layer 14 and the connections (not shown in FIG. 1) between these layers.
  • the code layer has only 64 neurons and thus a significantly smaller number of neurons than the input layer 11 and the output layer 14, i.e. H. the code layer 13 has a significantly smaller dimension than the input layer 11 and the output layer 14.
  • the artificial neural network of the auto-encoder 1 is therefore forced to learn to generate a code which is an efficiently compressed presentation of the input.
  • the encoder 3 has three hidden layers 12a, 12b and 12c between the input layer 11 and the code layer 13 and the decoder 5 has three hidden layers 12d, 12e, 12f between the code layer 13 and the output layer 14.
  • the hidden layers 12a, 12b, 12c, 12d, 12e, 12f are each designed as a convolutional layer.
  • the encoder 3 and the decoder 5 of the autoencoder 1 are designed symmetrically to one another.
  • the hidden layer 12a of the encoder 3 and the hidden layer 12f of the decoder 5 are designed as convolutional layers with eight 3 ⁇ 3 filters each, ie they have a depth of eight.
  • the hidden layer 12b of the encoder 3 and the hidden layer 12e of the decoder 5 are convolutional layers with sixteen 3 ⁇ 3 filters each (ie with a depth - 21
  • the hidden layer 12c of the encoder 3 and the hidden layer 12d of the decoder 5 are formed as convolutional layers each with two and thirty 3 x 3 filters (i.e. with a depth of 32). After each convolutional layer 12a, 12b, 12c of the encoder, a pooling layer (not shown in FIG. 1) is arranged.
  • the autoencoder 1 shown in FIG. 1 can be trained using a training data set and at least one error function so that the encoder 3 generates a code presenting the input from an input supplied to the autoencoder via the input layer 11, and the decoder 5 generates the code from the code generates an output output via the output layer 14 that is a reconstruction of the input.
  • FIG. 2 shows a schematic representation of an exemplary sequence of steps 101, 102, 103, 104 of the method according to the invention for training an autoencoder.
  • a training data set which includes both a plurality of training inputs of the first type and a plurality of training inputs of the second type.
  • a known expression of an attribute is assigned to each training input of the first type.
  • No known expression of the attribute is assigned to the training inputs of the second type.
  • FIG. 3 schematically shows a part of such a training data set 7, which contains a large number of training inputs of the first type 9a, 9b, 9c and a large number of training inputs of the second type 10.
  • the training inputs are designed as 28 ⁇ 28 pixel gray-scale images of the Fashion MNIST data set and each show an item of clothing.
  • the training data set 7 comprises both training inputs of the first type, ie training inputs that are provided with a label (labeled training inputs), and training inputs of the second type, ie training inputs that are not provided with a label (unlabeled training inputs).
  • the label represents the well-known expression of the Attribute that is assigned to the respective training input of the first type 9a, 9b, 9c.
  • the attribute is designed as a seasonal determination of the item of clothing shown on the respective image and can have the characteristics summer clothing, winter clothing or all-season clothing.
  • the attribute thus also describes belonging to a class, namely belonging to one of the three classes of summer clothing, winter clothing, all-year clothing.
  • the various inputs are assigned to the various classes, i.e. H. Assigned to the different versions of the attribute according to a simple assignment rule: sandals, dresses, shirts, t-shirts and tops are assigned to the summer clothing category, sweaters, coats and boots are assigned to the winter clothing class and pants, bags and sneakers are assigned to the category assigned to the year-round clothing class.
  • each training input of the first type of the training data set 7 is assigned a known expression of the attribute “seasonal determination”.
  • the training input 9a shown in FIG. 3, which shows trousers, is assigned the value all-season clothing.
  • the training input of the first type 9b, which shows a T-shirt, is assigned the value summer clothing.
  • the training input of the first type 9c, which shows a sweater, is assigned the expression winter clothing.
  • the “seasonal determination” attribute of the image data is an attribute whose expression cannot be readily recognized from the image data itself.
  • unsupervised learning such as is usually used to train an autoencoder
  • such an attribute is basically not taken into account and therefore has little or no influence on the distribution of the codes generated from the inputs in latent space.
  • the training inputs of the second type are therefore not provided with a label which indicates whether the respective training input is assigned to the class of summer clothing, winter clothing or year-round clothing.
  • the labeled training inputs of the first type thus only form a small part of the total number of training inputs in the training data set. When the invention was tested, it has been shown that such a small proportion of unlabeled training inputs is sufficient to achieve satisfactory results.
  • FIG. 2 also shows schematically that in step 102, for each training input of the first type, a target code assigned to the respective training input is determined as a function of the characteristic of the attribute assigned to the respective training input. For this purpose, a distance between the two characteristics is determined for each combination of two expressions of the attribute “seasonal determination”, which occurs in the entirety of the training inputs of the first type 9a, 9b, 9c of the training data set 7.
  • distance dsw between the versions summer clothing and winter clothing distance C / GS between the versions summer clothing and all-year clothing and distance C / GW between the versions winter clothing and all-year clothing.
  • distance C / GS distance between the versions summer clothing and all-year clothing
  • distance C / GW distance between the versions winter clothing and all-year clothing.
  • These distances can be chosen arbitrarily depending on the respective intended use of the method according to the invention.
  • the distances can in particular reflect the similarities and dissimilarities between the various manifestations of the attribute.
  • the index S denotes summer clothing
  • the index W denotes the variant winter clothing
  • the index G denotes the variant all year round clothing.
  • a distance between the training inputs is also assigned to each combination of training inputs of the first type of training data set 7.
  • each combination of training inputs of the first type which consists of a first training input with a first expression and a second training input with a second expression of the attribute, is assigned the distance between the first expression and the second expression as the distance between the training inputs.
  • the target codes are determined as a function of the distances thus assigned between the training inputs.
  • the distances between the training inputs 9a and 9b (pants and T-shirt, ie all-year-round clothing and summer clothes), between the training inputs 9a and 9c (trousers and sweater, ie all-season clothes and winter clothes) and between the training inputs 9b and 9c (T-shirt and pullover, ie summer clothing and winter clothing) each have the value 1.
  • the distances between the training inputs of the first type are determined so that the distances between the training inputs of the first type form a distance matrix D.
  • Each row and each column of the distance matrix D denotes a training input and the elements of the matrix D contain the Distances between training inputs.
  • the target codes are then obtained from this distance matrix D using multidimensional scaling (MDS), i.e. H. using a dimensional reduction method. This will be discussed in more detail later.
  • MDS multidimensional scaling
  • FIG. 2 also shows schematically that in a further step 103 of the method according to the invention for training the auto-encoder 1, an error function of the first type is provided.
  • x denotes an input and f ae (x) an output of the autoencoder generated from the input x.
  • the first error term is accordingly defined in this exemplary embodiment as the mean squared error (MSE) between the input x and the output f ae (x) generated from it. It thus describes a reconstruction error in the form of a discrepancy between a training input x and the output f ae (x) generated therefrom, if x is a training input.
  • MSE mean squared error
  • f enc (x) denotes the function with which the encoder generates a code from an input x, ie the projection function with which the encoder projects the input into the latent space.
  • f enc (x) denotes an actual code generated from a training input x, if x is a Training input is.
  • z denotes the target code assigned to the training input x. The determination of the target code z will be discussed in greater detail later.
  • the second error term Ls is thus defined as the mean square error between an actual code f enc (x) generated from a training input and the target code z assigned to the training input x.
  • the second error term thus describes a structuring error in the form of a discrepancy between an actual code generated from the training input and the target code assigned to the training input.
  • LSAE OT Z) YL S (f enc (x), z) + (1 - g) L AE (x, f ae (x)).
  • the first error term and the second error term are therefore linked to one another in the error function of the first type by an addition.
  • G denotes a weighting parameter with ge [0,1], which is used to determine within the error function of the first type between the first error term and the second error term - and thus between the reconstruction error and the structuring error - to be weighted.
  • the error function of the first type corresponds in principle to an error function as used in conventional training of an autoencoder. So that the error function of the first type describes not only the reconstruction error but also the structuring error, the weighting parameter y is therefore chosen such that 0 ⁇ g ⁇ 1 applies. A larger value of the weighting parameter / increases the significance of the structuring error within the error function.
  • step 104 shown in FIG. 2 the autoencoder is trained using the training data set 7 and the error function of the first type LSAE (X, Z).
  • the autoencoder is trained iteratively in a manner known per se, ie in several iterations, the value of the error function being determined after each iteration and the weights being determined by means of a backpropagation method of the artificial neural network.
  • every training input x of the training data set 7 is projected into the latent space, ie from every training input x a representing the input x, the code z is generated as
  • the totality of all vectors z is combined to form a matrix Z.
  • the target code z assigned to the respective training input x is determined in step 102 of the method according to the invention in this embodiment, as already explained above, as a function of the distance matrix D formed from the distances between the training inputs x.
  • a matrix Z generated which consists of the joined vectors z of the target codes assigned to the training inputs x.
  • MDS multidimensional scaling
  • a matrix B (ß £ y ⁇ ) is then calculated using the multidimensional scaling where ⁇ denotes the average of the respective column or row.
  • the target coordinates Z * result from scaling the matrix E k of the k largest eigenvectors of the matrix B with the associated eigenvalues l as
  • the matrix Z * thus contains the target coordinates in latent space determined by the multidimensional scaling.
  • An orthogonal matrix R is searched for which maps the matrix Z * determined as the result of the multidimensional scaling as closely as possible, ie with the least possible error, onto the matrix Z of the actual codes. It is sufficient to calculate an ideal rotation matrix around the coordinate origin, since the multidimensional scaling generates coordinates centered on the coordinate origin. The procedure for calculating this ideal rotation matrix R is explained in more detail below.
  • a new matrix S * is defined by copying S and setting all singular values other than 0 to the value 1.
  • the desired matrix Z which contains the target codes z for all training inputs of the first type in the form of vectors strung together, results from this as
  • the autoencoder is trained in step 104 shown in FIG. 2 using a training data set 7 that includes both training inputs of the first type and training inputs of the second type, as well as using the previously explained error function of the first type and an error function of the second type.
  • the error function of the second type describes the reconstruction error and corresponds to the above-explained first error term of the error function of the first type.
  • the training data set 7 can be divided into several batches for training the autoencoder and the training of the autoencoder can take place using the batches of the training data set 7.
  • the number of batches and epochs used for training the autoencoder is basically arbitrary. When testing the invention, it has been shown that a comparatively small number of iterations is sufficient to achieve satisfactory results.
  • FIGS. 4a and 4b show 2D projections of the codes generated by differently trained autoencoders in latent space.
  • the various markers cross, circle, triangle
  • codes that have been generated from entries of the same class are shown as points with the same marker (plus, point, cross).
  • FIG. 4a shows a projection of the codes in latent space as they were generated by an autoencoder trained in a conventional manner. It can be seen that the auto-encoder is not generating a satisfactory structuring of the codes in the latent space.
  • the codes belonging to the various classes of summer clothing, winter clothing and all-season clothing are not sufficiently separated from one another in the latent space, so that a classification based on the codes in the latent space is not reliably possible.
  • FIG. 4b shows, in a corresponding form of representation, a 2D projection of the codes in latent space, as they were generated by an autoencoder that has been trained with the inventive method for training an autoencoder.
  • the representation shows a clear structuring of the codes in the latent space, in which the codes according to different characteristics of the attribute, i.e. are separated from one another according to the different classes of summer clothing, winter clothing and year-round clothing.
  • the codes of the summer clothing class form a first group of codes 51 in the latent space
  • the codes of the winter clothing class form a second group 52 in the latent space
  • the codes of the all-season clothing class form a third group 53 in the latent space.
  • the codes grouped in this way are separated from one another according to the characteristics of the attribute, so that a classification according to the characteristics of the attribute is possible on the basis of the codes in the latent space.
  • FIG. 5 shows a schematic representation of an exemplary sequence of the method steps of a first embodiment of the method according to the invention for classifying data.
  • an autoencoder which is an encoder and a decoder, trained with a method for training an auto encoder of the type explained above.
  • an input data record is provided which comprises a number of inputs to be classified.
  • the inputs to be classified have the same form as the training inputs explained above, i.e. These are 28 x 28 pixel grayscale images from the Fashion MNIST dataset showing various items of clothing.
  • the attribute "seasonal determination" should serve as the classification criterion for the inputs to be classified, i.e. the entries to be classified should each be assigned to one of the classes summer clothing, winter clothing or all-year clothing.
  • a code is generated from each input of the input data set to be classified by means of the encoder of the auto encoder, which code represents the respective input.
  • each input to be classified is finally classified.
  • each entry is assigned a class (summer clothing, winter clothing or all-year clothing) depending on the code representing the respective entry.
  • step 114 of the method shown schematically in FIG. 5 takes place in this exemplary embodiment by a machine classification method, namely by a support vector machine (SVM).
  • SVM support vector machine
  • the Support Vector Machine determines a flyper level in the latent space, which acts as a class boundary between the classes of summer clothing, winter clothing and all-year clothing.
  • FIG. 6 shows a schematic representation of an exemplary sequence of Method steps of a second embodiment of the method according to the invention for classifying data.
  • the method steps 1 1 1 to 1 14 correspond to the first embodiment shown schematically in FIG. 5, so that reference can be made to the relevant statements.
  • a reliability indicator is determined for each classified input as a function of the code representing the respective input in latent space.
  • the reliability indicator shows the reliability of the classification decision made when assigning the respective class.
  • the reliability indicator is determined as the distance between the code representing the respective input, i.e. the coordinates in latent space that represent the respective input, to the flyper plane determined by the support vector machine, which forms the class boundary between the three different classes of summer clothing, winter clothing and all-season clothing. The greater this distance between the respective code and the hyperplane forming the class boundary, the higher the reliability of the classification decision.
  • a number of critical inputs is selected from the classified inputs depending on the reliability indicators determined for the inputs.
  • a predetermined proportion p of the classified inputs is selected, which consists of those classified inputs which have the lowest reliability of the classification decision (indicated by the reliability indicator).
  • step 1 17 the expression of the attribute is then determined manually for each of these selected critical inputs and assigned to the critical input.
  • a manual relabelling of the critical inputs is carried out in step 117.
  • the value of the attribute ie summer clothing, winter clothing or Year-round clothing, manually determined and assigned to the critical input.
  • FIG. 6 also shows that, after method step 117, method step 111 is repeated, ie. H. the autoencoder is trained using a method of the type explained above.
  • the critical inputs selected in step 116, to each of which a characteristic of the attribute was manually assigned in step 117, are used as training inputs of the first type.
  • the second embodiment of the method according to the invention for classifying data shown in FIG. 6 offers the advantage that the classification reliability can be considerably improved by manually relabelling the critical inputs.
  • FIG. 7 shows a schematic representation of an exemplary sequence of the method steps of a method according to the invention for generating an output.
  • a first method step 121 at least the decoder of an autoencoder which has been trained with a method for training an autoencoder of the type explained above is provided.
  • a code is determined in the form of a coordinate vector in latent space.
  • an output is generated from the code by means of the decoder of the autoencoder.
  • the training data set with which the autoencoder was trained does not include any image data of the in the exemplary embodiment shown in FIG Fashion MNIST data set.
  • the training data set includes 3D vector data generated with the Skinned Multi-Person Linear Model, which shows different human body shapes in a wide variety of postures.
  • the attribute relevant for training the car coder is the gender of the human body shown in the vector data, which can be male or female.
  • a 2D projection of the codes in latent space is shown in FIG with an autoencoder trained according to the invention from such inputs it has been generated.
  • a clear structuring of the codes associated with the different characteristics of the attribute can be seen, namely a first group of codes 68, which have been generated from the inputs with the expression of the attribute “male”, and a second group of codes
  • Codes 69 which have been generated from the expression of the attribute "female”. Also shown are the projections of the reference codes 61 and 62, which have been determined for the expression “male” and the expression “female” of the attribute. It can be seen that the reference codes 61, 62 form the respective center of the first group of codes 68 and the second group of codes 69 in the projection plane.
  • the code required for generating an output according to the invention is determined in this exemplary embodiment in that a transformation vector 64 is added to a reference code 63.
  • the transformation vector 64 corresponds to the difference vector between the different reference codes 62 and 61, which were determined when training the autoencoder for the different versions “female” and “male”.
  • the transformation vector 64 in this embodiment is thus a scaling of the difference vector with the scaling factor 1. Adding the transformation vector 64 to the reference code 63 results in the code 65 from which the decoder generates the output.
  • the outputs 66 and 67 are shown in FIG. 8 which the decoder of the autoencoder according to the invention generates as a reconstruction of the respective input from the code.
  • the output 66 is an output generated from the reference code 63
  • the output 67 is an output generated from the code 65. It can be seen that the output 66 shows a body with a typically male body shape, since the reference code 63 is in the area of the first
  • Group of codes 68 is arranged.
  • the method according to the invention for generating an output generates an output 67 in the embodiment shown in FIG. 8, which shows a human body with a typically female body shape. This is due to the fact that the code 65, from which the output 67 has been generated, is arranged in the area of the second group of codes 69.
  • FIG. 8 shows that outputs 66 and 67 are otherwise very similar. For example, the postures of the two bodies shown hardly differ from one another.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

L'invention concerne un procédé d'apprentissage d'un dispositif d'autoencodage, qui présente un dispositif d'encodage et un dispositif de décodage et qui peut être entraîné de manière telle que le dispositif d'encodage génère, à partir d'une saisie, un code représentant la saisie et que le dispositif de décodage génère une sortie à partir du code, qui est une reconstruction de la saisie. Selon l'invention, une structuration avantageuse du code dans l'espace latent est obtenue par les étapes suivantes consistant à : – utiliser un jeu de données d'apprentissage, qui comprend une multitude de saisies d'apprentissage d'un premier type, chaque saisie d'apprentissage du premier type étant associée à une expression connue d'au moins un attribut, – déterminer un code de consigne associé à la saisie d'apprentissage respective pour chaque saisie d'apprentissage du premier type en fonction de l'expression de l'attribut associé à la saisie d'apprentissage respective, – utiliser une fonction d'erreur d'un premier type, qui comprend un premier terme d'erreur et un deuxième terme d'erreur, le premier terme d'erreur décrivant une erreur de reconstruction sous forme d'un écart entre une saisie d'apprentissage et la sortie générée à partir de celle-ci et le deuxième terme d'erreur décrivant un écart entre un code instantané généré à partir d'une saisie d'apprentissage et le code de consigne associé à la saisie d'apprentissage, – entraîner le dispositif d'autoencodage à l'aide du jeu de données d'apprentissage et de la fonction d'erreur du premier type.
PCT/EP2019/056621 2019-03-15 2019-03-15 Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé WO2020187394A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/EP2019/056621 WO2020187394A1 (fr) 2019-03-15 2019-03-15 Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/EP2019/056621 WO2020187394A1 (fr) 2019-03-15 2019-03-15 Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé

Publications (1)

Publication Number Publication Date
WO2020187394A1 true WO2020187394A1 (fr) 2020-09-24

Family

ID=65904387

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2019/056621 WO2020187394A1 (fr) 2019-03-15 2019-03-15 Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé

Country Status (1)

Country Link
WO (1) WO2020187394A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112417919A (zh) * 2020-11-18 2021-02-26 珠海格力电器股份有限公司 一种防伪二维码的检测识别方法及系统、存储介质

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160321522A1 (en) * 2015-04-30 2016-11-03 Canon Kabushiki Kaisha Devices, systems, and methods for pairwise multi-task feature learning
US20170076224A1 (en) 2015-09-15 2017-03-16 International Business Machines Corporation Learning of classification model

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160321522A1 (en) * 2015-04-30 2016-11-03 Canon Kabushiki Kaisha Devices, systems, and methods for pairwise multi-task feature learning
US20170076224A1 (en) 2015-09-15 2017-03-16 International Business Machines Corporation Learning of classification model

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
BO YANG ET AL: "3D Object Reconstruction from a Single Depth View with Adversarial Learning", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 26 August 2017 (2017-08-26), XP080954656 *
SCHÖNEMANN, P. H.: "A Generalized Solution of the Orthogonal Procrustes Problem", PSYCHOMETRIKA, 1966
VAICIUKYNAS EVALDAS ET AL: "Learning Low-Dimensional Representation of Bivariate Histogram Data", IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, IEEE, PISCATAWAY, NJ, USA, vol. 19, no. 11, 1 November 2018 (2018-11-01), pages 3723 - 3735, XP011703112, ISSN: 1524-9050, [retrieved on 20181109], DOI: 10.1109/TITS.2018.2865103 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112417919A (zh) * 2020-11-18 2021-02-26 珠海格力电器股份有限公司 一种防伪二维码的检测识别方法及系统、存储介质

Similar Documents

Publication Publication Date Title
DE69838181T2 (de) Visualisierung und selbstorganisation multidimensionaler daten durch ausgeglichene orthogonale abbildung
DE69937897T2 (de) System und verfahren zur 4d rekonstruktion und darstellung
DE60318471T2 (de) Extraktion von wissen mittels eines objektbasierten semantischen netzes
EP1183619B1 (fr) Reseau semantique d'ordre n, operant en fonction d'une situation
DE112016004266T5 (de) Verfahren zur Gesichtsausrichtung
DE102005046747B3 (de) Verfahren zum rechnergestützten Lernen eines neuronalen Netzes und neuronales Netz
EP1238364A2 (fr) Procede de traitement de structures de donnees
DE10296704T5 (de) Fuzzy-Inferenznetzwerk zur Klassifizierung von hochdimensionalen Daten
DE112019007762T5 (de) Verfahren zum Trainieren eines generativen kontradiktorischen Netzes, modifiziertes Bildgenerierungsmodul und System zum Detektieren von Merkmalen in einem Bild
DE69330021T2 (de) Verbessertes System zur Mustererkennung für Sonar und andere Anwendungen
DE102019209644A1 (de) Verfahren zum Trainieren eines neuronalen Netzes
DE69223447T2 (de) Lernverfahren für neuronales Netzwerk und Klassifizieranlage zum Anwenden dieses Verfahrens
EP3719811A1 (fr) Consistance des identifications de données dans le traitement des images médicales permettant la classification de cellules
WO2021130340A1 (fr) Système et procédé d'assurance qualité de modèles basée sur des données
DE202017007534U1 (de) Multiskalige 3D-Textursynthese
WO2022037853A1 (fr) Procédé mis en œuvre par ordinateur pour analyser l'intérieur d'un véhicule
DE102021201124A1 (de) Trainieren von bildklassifizierernetzen
DE102018216078A1 (de) Verfahren und Vorrichtung zum Betreiben eines Steuerungssystems
EP1180258B1 (fr) Reconnaissance de formes par examen de caracteristiques complementaires apres traitement partiel
DE112014001697T5 (de) Bildverarbeitungsvorrichtung, Bildverarbeitungsprogramm und Betriebsverfahren für Bildverarbeitungsvorrichtung
WO2020187394A1 (fr) Procédé d'apprentissage d'un dispositif d'autoencodage et de classification de données ainsi que dispositif d'autoencodage et programme informatique associé
DE69802372T2 (de) Klassifizierungssystem und -verfahren mit N-Tuple- oder RAM-basiertem neuronalem Netzwerk
DE4121453C2 (de) Näherungsschlußfolgerungsvorrichtung
DE102020116013A1 (de) Verfahren zum Bereitstellen eines komprimierten künstlichen neuronalen Netzes mittels eines Autoencoders
DE102019213459A1 (de) Verfahren zum Komprimieren eines Neuronalen Netzes

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19712936

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19712936

Country of ref document: EP

Kind code of ref document: A1