KR20190015160A - Method for training artificial neural network - Google Patents

Abstract

Disclosed is a method for training at least one part of a neural network including a plurality of layers performed in a computing device according to one embodiment of the present disclosure. The method comprises the steps of: inputting training data including normal data and abnormal data to an input layer of a neural network; creating a histogram with feature values outputted at each of one or more hidden nodes of a hidden layer of the neural network for each training data to generate distribution of the feature value for each hidden node; calculating error of the distribution of each feature value and predetermined probability distribution; and selecting at least one hidden node of the hidden nodes of the hidden layer based on the error.

Description

{METHOD FOR TRAINING ARTIFICIAL NEURAL NETWORK}

This disclosure relates to artificial neural networks, and more specifically, to artificial neural network training methods and data classification through trained neural networks.

Pattern recognition is a field of machine learning, which means a discipline that recognizes the regularity of patterns and data. Pattern recognition techniques include supervised learning and unsupervised learning methods. The supervised learning method refers to a method in which the algorithm learns pattern recognition using data (referred to as "training" data) for which the result of pattern recognition has already been determined. Where each training data may be referred to as labeled data. An uncoordinated learning method means a method in which an algorithm finds a previously unknown pattern without labeled data.

A neural network may be used to implement the pattern recognition technique. The neural network consists of at least two nodes and links connecting these nodes. Each link may be weighted and the weight assigned to the link may be variable. The weight assigned to the link may be modified to be suitable for performing the pattern recognition that the neural network intends.

U.S. Patent No. 7,698,239 shows an example of such a neural network.

The present disclosure has been devised in response to the background art described above and is intended to provide a training method of an artificial neural network.

In accordance with one embodiment of the present disclosure for realizing the above-mentioned problem, a training method of at least a part of a neural network including a plurality of layers performed in a computing device is disclosed. The method includes inputting training data including normal data and abnormal data to an input layer of the neural network; Histogramizing feature values output from each of one or more hidden nodes of a hidden layer of the neural network for each training data to generate a distribution of feature values for the at least one hidden node; Calculating a distribution of each feature value and an error of a predetermined probability distribution; And selecting at least one hidden node of the at least one hidden node of the hidden layer based on the error.

Alternatively, the initial weights of the neural network including the plurality of layers may be determined at random.

Alternatively, the step of selecting at least one hidden node of the one or more hidden nodes of the hidden layer based on the error may include detecting a hidden node of the at least one hidden node of the hidden layer, the error being less than or equal to a predetermined value Selecting; And updating the weight of the unselected node to deactivate the unselected node except for the selected hidden node.

Alternatively, the neural network may include at least three hidden layers.

Alternatively, the training data may be unlabeled and the training method may be unsupervised.

Alternatively, the predetermined probability distribution may be a Weibull distribution whose parameters are randomly determined.

Alternatively, the method may further include normalizing a feature value output from the hidden layer.

Alternatively, the method may further comprise inputting the training data to the input layers of the plurality of neural networks, respectively.

Alternatively, the method may further comprise reconfiguring the neural network by ensemble one or more nodes selected from the plurality of neural networks.

A computer program stored on a computer readable storage medium is disclosed that includes a plurality of instructions executed by one or more processors of a computing device in accordance with another embodiment of the present disclosure. The computer program comprising instructions for causing training data including normal data and abnormal data to be entered into an input layer of a neural network; Instructions for generating histograms of feature values output at each of one or more hidden nodes of a hidden layer of the neural network for each training data to produce a distribution of feature values for the at least one hidden node; Calculate an error of a distribution of each feature value and a predetermined probability distribution; And selecting at least one hidden node of the one or more hidden nodes of the hidden layer based on the error.

A computing device is disclosed for training at least a portion of a neural network comprising a plurality of layers in accordance with another embodiment of the present disclosure. The computing device may include one or more processors; And a memory for storing instructions executable on the one or more processors,

Wherein the one or more processors are configured to input training data including normal data and abnormal data to an input layer of the neural network and to output feature values output from each of one or more hidden nodes of a hidden layer of the neural network for each training data to a histogram Generating a distribution of feature values for each of the at least one hidden node, calculating an error between a distribution of each feature value and a predetermined probability distribution, and determining, based on the error, At least one hidden node is selected.

The present disclosure can provide training methods for artificial neural networks.

1 is a schematic diagram illustrating a portion of an artificial neural network in accordance with one embodiment of the present disclosure;
Figure 2 shows a block diagram of a computing device in accordance with one embodiment of the present disclosure.
FIG. 3A is a diagram showing a Weibull distribution. FIG.
FIG. 3B is a diagram showing a comparison of the distribution of a specific bayonet distribution and a feature value.
4 is a diagram illustrating histograms of features extracted from input data according to one embodiment of the present disclosure to compare with the Bayesian distribution.
5 illustrates selecting at least some nodes of a hidden layer in accordance with one embodiment of the present disclosure.
6 is a diagram illustrating ensemble selected nodes in a plurality of neural networks in accordance with one embodiment of the present disclosure.
7 is a block diagram of a computing device according to one embodiment of the present disclosure.

Various embodiments are now described with reference to the drawings, wherein like reference numerals are used throughout the drawings to refer to like elements. In this specification, various explanations are given in order to provide an understanding of the present disclosure. It will be apparent, however, that such embodiments may be practiced without these specific details. In other instances, well-known structures and devices are provided in block diagram form in order to facilitate describing the embodiments.

The terms "component," "module," system, "and the like, as used herein, refer to a computer-related entity, hardware, firmware, software, combination of software and hardware, or execution of software. For example, a component may be, but is not limited to, a process, a processor, an object, an execution thread, a program, and / or a computer running on a processor. For example, both an application running on a computing device and a computing device may be a component. One or more components may reside within a processor and / or thread of execution, one component may be localized within one computer, or it may be distributed between two or more computers. Further, such components may execute from various computer readable media having various data structures stored therein. The components may communicate, for example, via a network (e.g., a local system, data and / or signals from one component interacting with another component in a distributed system, and / or signals with one or more data packets) And / or < / RTI > transmitted data).

The description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present disclosure should not be construed as limited to the embodiments set forth herein, but is to be accorded the widest scope consistent with the principles and novel features presented herein.

1 is a schematic diagram illustrating a portion of an artificial neural network in accordance with one embodiment of the present disclosure;

Throughout this specification, neural networks, network functions, and neural networks can be used interchangeably. A neural network may consist of a collection of interconnected computational units, which may be referred to generally as " nodes ". These " nodes " may also be referred to as " neurons ". The neural network is configured to include at least two nodes. The nodes (or neurons) that make up the neural networks may be interconnected by one or more " links ".

Within a neural network, two or more nodes connected through a link may form a relationship of input and output nodes relatively. The concepts of the input node and the output node are relative, and any node in the output node relationship with respect to one node can be in the input node relationship with the other node, and vice versa. As described above, the input node-to-output node relationship can be generated around the link. One or more output nodes may be connected to one input node via a link, and vice versa.

In an input node and an output node relationship connected through one link, the output node can be determined based on data input to the input node. Here, a node interconnecting the input node and the output node may have a weight. The weights may be variable and may be varied by the user or algorithm to perform the desired function of the neural network. For example, if one or more input nodes to one output node are interconnected by respective links, then the output node is set to the values input to the input nodes connected to the output node and to the links corresponding to the respective input nodes The output node value can be determined based on the weight.

As described above, a neural network is formed by interconnecting two or more nodes through one or more links to form an input node and an output node relationship within a neural network. The nature of the neural network can be determined according to the number of nodes and links in the neural network, the association between the nodes and links, and the value of the weight assigned to each of the links. For example, if there are the same number of nodes and links, and there are two neural networks with different weight values between the links, then the two neural networks may be perceived as being different from each other.

As shown in FIG. 1, the neural network may be configured to include two or more nodes. Some of the nodes that make up the neural network can construct a layer based on the distances from the original input node. For example, n layers can be configured. The distance from the original input node may be defined by the minimum number of links that must go through to reach that node from the original input node. However, the definition of such a layer is arbitrary for explanation, and the degree of the layer in the neural network can be defined in a manner different from that described above. For example, the layer of nodes may be defined by distance from the final output node.

The first input node may refer to one or more nodes in the neural network to which data is directly input without going through a link in relation to other nodes. Or may refer to nodes in a neural network that do not have other input nodes connected by a link in the relationship between the nodes based on the link. Similarly, the final output node may refer to one or more nodes that do not have an output node, in relation to other ones of the nodes in the neural network. Also, the hidden node may refer to nodes constituting a neural network other than the first input node and the last output node. 1, the output node is omitted. The neural network according to an embodiment of the present disclosure may be a neural network in which the node of the input layer may be larger than the node of the hidden layer closer to the output layer and the number of nodes decreases as the input layer progresses to the hidden layer.

Figure 2 shows a block diagram of a computing device in accordance with one embodiment of the present disclosure.

The computing device 100 in accordance with one embodiment of the present disclosure may include a processor 110, a GPU 120, and a memory 130. The block diagram shown in FIG. 1 is a simplified representation of a computing device, and the present disclosure is not so limited and may include additional components necessary for driving.

The processor 110 may be comprised of one or more and may include a central processing unit (CPU) of the computing device. The processor 110 may read a computer program stored in the memory 130 to perform a training method of an artificial neural network (ANN) according to an embodiment of the present disclosure and a data classification method using a trained neural network . In accordance with one embodiment of the present disclosure, the processor 110 or the GPU 120 may perform calculations for training of a neural network. Although a processor and a GPU are shown separately in Fig. 2, in one embodiment of the present disclosure, a processor may include both a CPU and a GPU, and may include any processing devices for processing a neural network. The method according to one embodiment of the present disclosure may be performed by a processor.

A graphics processing unit (GPU) 120 is a graphics processing unit (GPU) that processes the input data for learning in deep learning (DN), extracts features from input data, computes the error and weight of the neural network using backpropagation Updating can be performed for learning of the neural network. Also, a computer program executed on a computing device according to an embodiment of the present disclosure may be a GPU-executable program. Also not shown in FIG. 1, but in accordance with one embodiment of the present disclosure, a computing device may include a tensor processing unit (TPU).

The memory 130 may store a computer program for performing the artificial neural network training method and data classification method according to one embodiment of the present disclosure and the stored computer program may be read by the processor 110, the GPU 120, Can be driven.

Hereinafter, a training method and a data classification method of an artificial neural network according to an embodiment of the present disclosure will be described.

Hereinafter, terms used in this specification will be defined.

A feature may correspond to at least one character of the input data. The characteristic may be, for example, a gray scale histogram of the image if the input data is an image, a color distribution, a color temperature, an edge, an edge shape, a combination of these, and the like, and the above- Do not. The feature may be computed by at least a portion of a network function to correspond to at least one characteristic of the input data quantified. For example, it is assumed that there is a characteristic indicating that the product is defective in the inspection of the defective product (for example, in the case of the defect inspection of the leather, the imprint of the leather, The feature may be computed based on the input data by a network function and may correspond to the property. The description of the characteristics described above is merely an example, and the present disclosure is not limited thereto. The feature in the neural network may correspond to a node determined by a constant sub-structure of the neural network based on the input data. For example, there may be one or more nodes corresponding to characteristics representing the impression of leather. A feature represents a characteristic of the input data and may include a part of the input data, a summary of the input data, and a meaning extracted from the input data. For example, it may be a process up to a particular node of a particular layer that processes input data. For example, the feature may be one or more values calculated by the neural network connection, weight, bias, etc. connected to one of the hidden layers of the neural network.

The characteristic is a characteristic part of the data, and in the case of image data for inspection of a defective product as in the above-described example, the part that can recognize that the product is defective in the image data may be a characteristic. That is, the image data of the product may be characteristics such as imprinting or dyeing to recognize that the product is defective. This characteristic can be recognized in the neural network corresponding to one or more features. Characteristics and features can have a one-to-one correspondence, one-to-many (many), many-to-one or many-to-many correspondence relationship. That is, if the image data of the product is such that the product is bad, the feature may be a value based on such a characteristic of the data. The one or more features may be related to one or more characteristics. In the above example, the feature may be a sub-structure of a neural network that allows the feature to be extracted from the image data. In addition, the characteristic extracted through the feature can be not only visually confirmed by the user or the like as in the above-mentioned example, but also includes an abstract characteristic capable of discriminating the data. The above-described characteristic is merely an example, and the present disclosure is not limited thereto.

The feature map may be an image representing a feature on an image and representing a feature by location on the image.

The feature value corresponds to the above-described feature, and may mean an output value obtained by processing input data with a neural network. The feature value may be the result of processing the input data into a feature. For example, the output of the particular node of the input data may be a feature value if the neural network connection to the particular node included in the hidden layer of the neural network implies a feature. The feature value can be calculated by the neural network based on each input data. The feature value can have any dimension. The feature value may be an output value obtained by processing each input data with a neural network. Feature values may exist corresponding to each of the input data. For example, a connection path, a weight, a bias, and a calculation process to a node 1 of the hidden layer 3 correspond to a feature, and an output value output from the corresponding node with respect to input data may be a feature value.

The histogram is a graphical representation of the form of the frequency distribution, which in the present disclosure can mean graphically representing the shape of the distribution of feature values. For example, the histogram may represent the form of the frequency distribution as a columnar graph, but the disclosure is not limited thereto. The feature value is a discrete value. A sample having each feature value, the number of input data, may be represented by a histogram. The feature values for each input data may be histogrammed and represented as a distribution of feature values.

A computing device according to one embodiment of the present disclosure processes training data in a neural network to histogram the output of feature values to generate a distribution of feature values, compares the distribution of feature values with a predetermined probability distribution, Based on the error, some nodes of the neural network can be selected to pre-train the neural network.

The neural network of the present disclosure may be used for classification of data. For example, the neural network of the present disclosure can be used for anomaly detection. More specifically, the neural network of the present disclosure can be used to classify defective and normal products in the production process. The training data of the present disclosure may include image data, the normal data may include an image of a normal product, and the abnormal data may include an image of a defective product. The scope of the present disclosure is not limited to the processing of image data.

According to the data classification using the neural network of the present disclosure, when the distribution of the feature values extracted from the input data has an error below a predetermined probability distribution and a predetermined threshold value, the feature that outputs the feature value is suitable for classifying the input data Can be interpreted as a feature. In general factories, the distribution of stationary and defective products follows a natural probability distribution. Thus, if the distribution of feature values is similar to this natural probability distribution, then the feature is likely to be related to an important characteristic for data classification. Thus, the computing device 100 may activate or deactivate a portion of the neural network to use the feature for classification. According to one embodiment of the present disclosure, computing device 100 may choose to use the feature for classification if the distribution of feature values follows a particular probability distribution. Hereinafter, a comparison between the distribution of the probability distribution and the feature value will be described.

FIG. 3A is a view showing the bay fire distribution. FIG.

Bayfire distribution is one of the continuous probability distributions and can represent various probability density function types depending on parameters. Bayfire distributions can take the form of various types of probability density functions depending on the variation of the parameters lambda and k. The Bayfire distribution can exhibit most probability distributions as the parameters change, and the Bayfire distribution shown in FIG. 3a is only an example, and the probability distribution according to one embodiment of the present disclosure may include a multi-dimensional probability distribution . In another embodiment, the probability distribution according to one embodiment of the present data may be based on historical statistical values of the analysis data.

If the distribution of the feature values has a shape similar to that of the graph of y = a rather than the shape of the probability distribution, the input data is output as the respective feature values, so if the distribution of the feature values has the shape described above, Is a feature that extracts the characteristics of each input data. However, even if the characteristics of input data are extracted, it does not help classification of input data.

However, if the distribution of feature values extracted from the input data has a shape of a predetermined probability distribution, the feature extracts characteristics of a part or a plurality of characteristics of the input data. Can be an important feature.

If the distribution of the feature values follows a probability distribution as shown in FIG. 3B, the feature may be a feature extracting a characteristic that only a part of the input data has. The characteristic that only a part of such input data has can be an important characteristic for classification of input data. Thus, if the distribution of feature values is similar to a particular probability distribution, the feature may be an important feature for classification of the input data, and the computing device 100, according to one embodiment of the present disclosure, The node can be activated or deactivated.

Accordingly, when the distribution of the feature values output from the hidden node of the hidden layer is similar to the predetermined probability distribution function, the neural network according to an embodiment of the present disclosure is characterized in that the feature (i.e., corresponding node of the neural network) feature, and the node outputting the feature value may be a node outputting the important feature in the data classification. The determination of the similarity between the distribution of the feature values and the predetermined probability distribution function will be described below. Therefore, it is possible to reconfigure the neural network by activating only nodes outputting such important features, or to extract nodes that output important features from a plurality of neural networks. The neural network reconstructed in this way can have high accuracy in classification of data, and it is possible to improve the training efficiency by reducing the number of iterations and calculation amount of training by configuring the neural network to extract important features before training.

Because the predetermined probability distribution is a continuous function and the histogram is discrete, the computing device 100 may determine an error in distribution of the predetermined probability distribution and the feature value according to one embodiment of the present disclosure. FIG. 3B is an exemplary diagram for judging the similarity of the distribution of the predetermined probability distribution and the feature value. FIG. In the example of FIG. 3B, the curve representing the continuous probability distribution may represent a Bayesian distribution with parameters lambda = 1, k = 0.5. In the example of FIG. 3B, the bar graph shaped portion may represent a distribution of feature values. In the example of FIG. 3B, the x-axis is the feature value and the y-axis represents the number of specimen (i.e., input data) by feature value. The x and y axes may be normalized (i.e., the probability distribution curve may have a value of 0 < y < 1) for comparison with a probability distribution curve. That is, computing device 100 may facilitate comparison with a predetermined probability distribution by normalizing values corresponding to each axis of the distribution of feature values. In the example of FIG. 3B, the feature has the largest number of input data having the x1 value, and the number of input data having the corresponding value decreases as the value moves from the x2 value to the xn value. The computing device 100 in accordance with one embodiment of the present disclosure may determine a distribution of feature values and an error of a predetermined probability distribution. 3B, the computing device 100 compares the number of input data (or the normalized number of input data numbers) of each of the feature values with a predetermined probability distribution value at each of the points . For example, in the example of FIG. 3B, the computing device 100 compares (x1, y1) and one or more points of a predetermined probability distribution graph (i.e., y values on the graph corresponding to x1) And the distribution of feature values. For example, the variance between the value of the normalized histogram and the corresponding probability distribution value can be used as a comparison criterion.

In another embodiment, the computing device 100 may determine a degree of similarity between a distribution of feature values and a predetermined probability distribution based on whether the distribution of feature values exhibits a trend similar to a particular probability distribution. For example, the difference between the derivative of the normalized histogram and the corresponding derivative of the probability distribution can be used as a comparison criterion. Y2 &gt; y3 &gt; y4 &gt; y5 in the case where the distribution of the feature value shows a tendency to decrease, i.e., x1 < x2 &lt; x3 &lt; x4 & ) Can be judged to be similar to the Bayesian distribution where the distribution of the feature values is the parameter λ = 1, k = 0.5. That is, in accordance with one embodiment of the present disclosure, the computing device 100 may determine a probability distribution based on whether a feature value in a distribution of feature values and a number of input data pertaining to each feature value exhibit a similar tendency to a predetermined probability distribution It is possible to judge the similarity and the error of the distribution of the feature values. If the error is below a predetermined threshold, the computing device 100 may determine that the distribution of the predetermined probability distribution and the feature value are similar. The above description is only an example, and the probability distribution of the present disclosure, the shape of the feature value distribution and the similarity determination method are not limited thereto.

The computing device 100 according to one embodiment of the present disclosure may input training data, including normal and abnormal data, to the input layer of the neural network for training of the neural network. Training data according to one embodiment of the present disclosure may be unlabeled data.

The computing device 100 according to an exemplary embodiment of the present invention calculates training data input from an input layer and outputs an output value. The output value of the input layer is transferred to a hidden layer, Can be output. 4 is an exemplary diagram illustrating a portion of an artificial neural network in accordance with one embodiment of the present disclosure; The neural network according to one embodiment of the present disclosure may include at least three hidden layers. The number of layers and the number of nodes of the neural network shown in FIG. 4 are merely examples, and the present disclosure is not limited thereto. An artificial neural network according to an embodiment of the present disclosure may be a convolution neural network (CNN), an auto encoder, a feedforward neural network, a radial basis function network, A kohonen self-organizing network, a recurrent neural network (RNN), etc. The present disclosure is not limited thereto and may include any neural network.

The computing device 100 according to one embodiment of the present disclosure may generate a distribution of feature values by histogramming the feature values output at the hidden node of the hidden layer of the neural network for each training data. The hidden layer may be a layer before the output layer of the neural network. Referring to Figure 4 hidden layer A set of values, calculated by the connection path, weight, of the node up to n1 node of 3 may be a feature . The feature value may include any characteristic, such as contrast, brightness, color distribution, image edge, image histogram, etc., of the image as a result of processing the training data at each node of each layer, But can also include abstract properties. In the example of FIG. 4, the hidden layer 3 can extract three features (the calculation process up to n1, n2, and n3 nodes), and each node of the hidden layer 3 can output the feature value for each training data have.

In the example of FIG. 4, the first feature value, which is the output of the n1 node of the hidden layer 3 (the calculation result of the input data for the first feature), may be histogrammed and output as the first feature value distribution 310. [ The computing device 100 may generate a first feature value distribution 310 by histogramming the output values of each of the input data at node n1.

Here, the first feature value distribution 310 represents the feature value (output of n1) for each input data by histogram, the x-axis represents the feature value, and the y-axis represents the number of samples in the feature value. One feature value may be outputted for one input data, and one output value of the output may be one sample in the histogram. The first feature value distribution is a histogram of feature values for each of a plurality of input data. Thus, the output for each of the input data at the n2 node of the hidden layer 3 may be the second feature value distribution 330 and the output for each of the input data at the n3 node of the hidden layer 3 may be the third feature value distribution 350 ).

The computing device 100 in accordance with one embodiment of the present disclosure may calculate a distribution of feature values and an error of a predetermined probability distribution. The predetermined probability distribution may comprise a Bayesian distribution. Bay fire distribution 200 is one of the continuous probability distributions as shown in FIG. 3A, and can represent various types of probability density functions according to parameters. Bayfire distributions can take the form of various types of probability density functions depending on the variation of the parameters lambda and k. As shown in FIG. 3A, various types of distribution of probability density function can be obtained according to each parameter. Where the parameters of the bay fire distribution can be determined at random and can be determined to have an appropriate shape according to generally known defect rates. The Bayfire distribution described above is only an example, and the present disclosure may include an n-dimensional probability distribution.

The computing device 100 may calculate the error of the feature value distribution 310,330, 350 and the predetermined probability distribution 200. [ The computing device 100 may compute the error by comparing the distribution of feature values 310, 330, 350 with a predetermined bayonet distribution. In the example of FIG. 4, the first feature value distribution 310 shows a histogram similar to the shape of the graph of the comparative probability distribution, where the feature values are not uniformly distributed over several values, but are concentrated on some values. More specifically, when looking at the shape of the first feature value distribution 310, most of the feature values are concentrated in some areas, and there are almost no feature values at the end of the x-axis value Seems to be. Therefore, the first feature for extracting the first feature value from the viewpoint of this shape is determined to be an important feature suitable for dividing the characteristics possessed by most of the input data and the features not possessed by the most input data. There is a feature value in which there are many specimens (x1 point with the highest y-axis value at reference numeral 310 in Fig. 4) and a feature value with a small specimen (see Fig. 4) in the shape of the first feature value distribution 310 Number 310 x2 point with low y-axis value). It can be seen that the feature value for each of the input data corresponding to the n1 feature has the largest value of x1 and the least value of x2. This indicates that the number of input data for which the feature value x1 is output and the number of input data for which the feature value x2 is output are small. Thus, the n1 feature can be a feature that extracts x1 as a feature value from most input data for a large number of input data (typically, more normal data than abnormal data), and extracts x2 as a feature value from some input data have. The first feature value distribution 310 extracted through the n1 feature may have a bay fill distribution shape with a concentrated number of samples for some feature values (a shape similar to a solid line in bay fill distribution 200). The computing device 100 may compute an error between the first feature value distribution 310 and a predetermined bay dominance (solid line 200), and if the error is less than or equal to the predetermined value, the computing device 100 may calculate the first feature value (In this example, node n1 in the example of FIG. 4).

Similarly, a second feature value for each input data may be output at the n2 node of the hidden layer 3, and the computing device 100 may generate the second feature value distribution 330. [ The second feature value distribution 330 has a shape in which the sample is not concentrated on a specific feature value but the number of samples is uniformly distributed as a whole. In this case, the second feature can be assumed to be a feature extracting characteristic of most images. The second feature value distribution 330 has no feature value with a significantly larger sample than the other feature values, which means that the feature extracted from the second feature is a characteristic that the input data may have in general. The computing device 100 may calculate the error of the second feature value distribution 330 and the predetermined bay fade distribution. In this case, the error of the second feature value distribution 330 and the predetermined bayonet distribution may be greater than the error of the first feature value distribution 330 and the predetermined bayonet distribution. Accordingly, the computing device 100 may determine that the second feature is an unnecessary feature in input data classification if the error between the second feature value distribution 330 and the predetermined bay dominance is greater than a predetermined value (i.e., a threshold) . Thus, the computing device 100 may deactivate the node (in this example, the node n2 in this example) that outputs the second feature value.

In the same manner, the distribution of the feature values output from the node n3 of the hidden layer 3 also has a shape in which the sample is slightly concentrated in the specific feature value. The computing device 100 may calculate the error of the third feature value distribution 350 and the predetermined bay dominance 200 to activate the n3 node outputting the third feature value if the error is below a predetermined value .

Activation and deactivation of the nodes described above can be performed by adjusting weights connecting the node and the next node. For example, the computing device 100 may adjust the connection weight between the node of the output layer connected to node n1 and node n1 to one so that the output of the selected node n1 is fully delivered to the output layer. In addition, for example, the computing device 100 may adjust the connection weight between the node of the output layer connected to the n2 node and the node n2 to zero so that the output of the deactivated n2 node is not transmitted to the output layer. The above-described connection weight is only an example, and the present disclosure is not limited thereto. The computing device 100 may set appropriate connection weights so that the output of the deactivated node is not delivered to the next layer and may set an appropriate weight to allow the output of the activated node to be delivered to the next layer.

The computing device 100 according to one embodiment of the present disclosure may normalize the feature value output from the hidden layer. In the example of FIG. 4, the range of feature values (x-axis of 310,330, 350) in each feature value distribution may be different from each other. Thus, the computing device 100 may normalize the feature value to match the x-axis of the feature value distribution to facilitate comparison of each feature value distribution with a predetermined probability distribution.

FIG. 5 is a diagram showing a neural network in which nodes n1 and n3 of the hidden layer 3 are selected in the example of FIG. 4 and n2 is inactivated. The computing device 100 may select the nodes n1 and n3 and adjust the weight of each node so that the output of the corresponding node is transferred to the next layer. The computing device 100 may deactivate the node n2 and adjust the weight of the n2 node such that the output of the node is not delivered to the next layer.

6 is an illustration showing that a computing device forms a neural network by ensemble a plurality of neural networks.

Comparison of the distribution of feature values of the present disclosure described above with a predetermined probability distribution and thus deactivation of a portion of the neural network can be performed for each of a plurality of neural networks.

For example, in the example of FIG. 6, the nodes n1 and n3 of the hidden layer 3 are activated and the node n2 is deactivated in the first neural network 400, and the node n4 of the hidden layer 3 is activated in the second neural network 410 Other nodes were disabled.

In this case, the computing device 100 may configure the third neural network 430 by ensemble the first neural network 400 and the second neural network 410.

For example, node n4 of the second neural network 410 receives input from nodes 4, 5, 6, and 7 of hidden layer 2, nodes 4, 5, 6, and 7 of hidden layer 2 receive node 6 , 7, 8, 9, and 10, respectively. This connection relationship is also maintained in the third neural network 430.

In the case of applying the pre-training method according to an embodiment of the present disclosure in parallel to a plurality of neural networks, the network structure can be simplified in the pre-training step, the amount of calculation can be reduced, sensitivity testing is easy, (For example, in the case of performing the preliminary training method of the present disclosure through two integrated networks, for example, in the case of FIG. 6, hidden layer 1 is extended from 5 nodes to 10 nodes , Hidden layer 2 increases from four nodes to eight nodes, increasing the overall amount of computation, and sensitivity testing costs).

The ensemble scheme described in FIG. 6 is merely an example, and the scope of the present disclosure may include various ensemble techniques.

Normal data obtained by photographing a normal product and abnormal data obtained by photographing an abnormal product generally have a problem that a large number of normal data is acquired than abnormal data and it is difficult to acquire a sufficient number of abnormal data. Abnormal data is difficult to obtain sufficient numbers for learning and various data augmentation techniques are used to increase the data to obtain a sufficient number of abnormal data for learning. However, in accordance with one embodiment of the present disclosure, if the distribution of feature values output from the neural network precondition the neural network to use only features (significant features) similar to the predetermined probability distribution, Important features (ie, important nodes) can be configured to improve the training efficiency by reducing the number of iterations and calculation amount of training. In this case, unlabeled training data can be trained, so that more data can be processed by non-supervised training method, which can improve accuracy.

The neural network training method according to an embodiment of the present disclosure can be performed before repeated learning of a neural network such as back propagation. A neural network training method according to an embodiment of the present disclosure compares a distribution of feature values extracted at each node of a neural network with a predetermined probability distribution to determine whether or not the distribution of feature values is a node that extracts only feature values on a probability distribution Only the node extracting only the important features can be included in the neural network. Using the neural network training method according to an embodiment of the present disclosure, it is possible to determine a neural network composed of only nodes extracting only important features before iterative learning, thereby improving the efficiency in the later iterative learning and improving the classification accuracy .

7 is a block diagram of a computing device according to one embodiment of the present disclosure.

Figure 7 illustrates a simplified, general schematic diagram of an exemplary computing environment in which embodiments of the present disclosure may be implemented.

Although the present disclosure has been described above generally in terms of computer-executable instructions that may be executed on one or more computers, those skilled in the art will appreciate that the disclosure may be combined with other program modules and / will be.

Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Those skilled in the art will also appreciate that the methods of the present disclosure may be practiced with other computer systems, including single-processor or multiprocessor computer systems, minicomputers, mainframe computers, as well as personal computers, handheld computing devices, microprocessor-based or programmable consumer electronics, And may operate in conjunction with one or more associated devices).

The described embodiments of the present disclosure may also be practiced in distributed computing environments where certain tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

Computers typically include a variety of computer readable media. Any medium accessible by a computer may be a computer-readable medium, which may include volatile and non-volatile media, transitory and non-transitory media, removable and non-removable media, Removable media. By way of example, and not limitation, computer readable media can comprise computer readable storage media and computer readable transmission media. Computer-readable storage media includes both volatile and non-volatile media, both temporary and non-volatile media, both removable and non-removable, implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data Media. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital video disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, Or any other medium which can be accessed by a computer and used to store the desired information.

Computer readable transmission media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism, It includes all information delivery media. The term modulated data signal refers to a signal that has one or more of its characteristics set or changed to encode information in the signal. By way of example, and not limitation, computer readable transmission media include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, or other wireless media. Combinations of any of the above described media are also intended to be included within the scope of computer readable transmission media.

There is shown an exemplary environment 1100 that implements various aspects of the present disclosure including a computer 1102 and a computer 1102 that includes a processing unit 1104, a system memory 1106, and a system bus 1108 do. The system bus 1108 couples system components, including but not limited to, system memory 1106 to the processing unit 1104. The processing unit 1104 may be any of a variety of commercially available processors. Dual processors and other multiprocessor architectures may also be used as the processing unit 1104.

The system bus 1108 may be any of several types of bus structures that may additionally be interconnected to a local bus using any of the memory bus, peripheral bus, and various commercial bus architectures. The system memory 1106 includes read only memory (ROM) 1110 and random access memory (RAM) The basic input / output system (BIOS) is stored in a non-volatile memory 1110, such as a ROM, EPROM, EEPROM or the like, which is a basic (non-volatile) memory device that aids in transferring information between components within the computer 1102 Routine. The RAM 1112 may also include a high speed RAM such as static RAM for caching data.

The computer 1102 may also be an internal hard disk drive (HDD) 1114 (e.g., EIDE, SATA) - this internal hard disk drive 1114 may also be configured for external use within a suitable chassis , A magnetic floppy disk drive (FDD) 1116 (e.g., for reading from or writing to a removable diskette 1118), and an optical disk drive 1120 (e.g., a CD-ROM For reading disc 1122 or reading from or writing to other high capacity optical media such as DVD). The hard disk drive 1114, magnetic disk drive 1116 and optical disk drive 1120 are connected to the system bus 1108 by a hard disk drive interface 1124, a magnetic disk drive interface 1126 and an optical drive interface 1128, respectively. . The interface 1124 for external drive implementation includes at least one or both of USB (Universal Serial Bus) and IEEE 1394 interface technologies.

These drives and their associated computer-readable media provide non-volatile storage of data, data structures, computer-executable instructions, and the like. In the case of computer 1102, the drives and media correspond to storing any data in a suitable digital format. While the above description of computer readable media refers to HDDs, removable magnetic disks, and removable optical media such as CDs or DVDs, those skilled in the art will appreciate that other types of storage devices, such as zip drives, magnetic cassettes, flash memory cards, Or the like may also be used in the exemplary operating environment and any such medium may include computer-executable instructions for carrying out the methods of the present disclosure.

A number of program modules may be stored in the drive and RAM 1112, including an operating system 1130, one or more application programs 1132, other program modules 1134, and program data 1136. All or a portion of the operating system, applications, modules, and / or data may also be cached in the RAM 1112. It will be appreciated that the disclosure may be implemented in a variety of commercially available operating systems or combinations of operating systems.

A user may enter commands and information into the computer 1102 via one or more wired / wireless input devices, such as a keyboard 1138 and a pointing device such as a mouse 1140. [ Other input devices (not shown) may include a microphone, IR remote control, joystick, game pad, stylus pen, touch screen, and so on. These and other input devices are often connected to the processing unit 1104 via an input device interface 1142 that is coupled to the system bus 1108, but may be a parallel port, an IEEE 1394 serial port, a game port, a USB port, an IR interface, &Lt; / RTI &gt; and so forth.

A monitor 1144 or other type of display device is also connected to the system bus 1108 via an interface, such as a video adapter 1146, In addition to the monitor 1144, the computer typically includes other peripheral output devices (not shown) such as speakers, printers,

Computer 1102 may operate in a networked environment using logical connections to one or more remote computers, such as remote computer (s) 1148, via wired and / or wireless communication. The remote computer (s) 1148 may be a workstation, a computing device computer, a router, a personal computer, a portable computer, a microprocessor-based entertainment device, a peer device or other conventional network node, But for the sake of simplicity, only memory storage device 1150 is shown. The logical connections depicted include a wired / wireless connection to a local area network (LAN) 1152 and / or a larger network, e.g., a wide area network (WAN) These LAN and WAN networking environments are commonplace in offices and corporations and facilitate enterprise-wide computer networks such as intranets, all of which can be connected to computer networks worldwide, for example the Internet.

When used in a LAN networking environment, the computer 1102 is connected to the local network 1152 via a wired and / or wireless communication network interface or adapter 1156. [ The adapter 1156 may facilitate wired or wireless communication to the LAN 1152 and the LAN 1152 also includes a wireless access point installed therein to communicate with the wireless adapter 1156. [ When used in a WAN networking environment, the computer 1102 may include a modem 1158, or may be connected to a communications computing device on the WAN 1154, or to establish communications over the WAN 1154 And other means. A modem 1158, which may be an internal or external and a wired or wireless device, is coupled to the system bus 1108 via a serial port interface 1142. In a networked environment, program modules described for the computer 1102, or portions thereof, may be stored in the remote memory / storage device 1150. It will be appreciated that the network connections shown are exemplary and other means of establishing a communication link between the computers may be used.

The computer 1102 may be any wireless device or entity that is deployed and operable in wireless communication, such as a printer, a scanner, a desktop and / or portable computer, a portable data assistant (PDA) Any equipment or place, and communication with the telephone. This includes at least Wi-Fi and Bluetooth wireless technology. Thus, the communication may be a predefined structure, such as in a conventional network, or simply an ad hoc communication between at least two devices.

Wi-Fi (Wireless Fidelity) allows you to connect to the Internet without wires. Wi-Fi is a wireless technology such as a cell phone that allows such devices, e.g., computers, to transmit and receive data indoors and outdoors, i. E. Anywhere within the coverage area of a base station. Wi-Fi networks use a wireless technology called IEEE 802.11 (a, b, g, etc.) to provide a secure, reliable, and high-speed wireless connection. Wi-Fi can be used to connect computers to each other, the Internet, and a wired network (using IEEE 802.3 or Ethernet). The Wi-Fi network may operate in unlicensed 2.4 and 5 GHz wireless bands, for example, at 11 Mbps (802.11a) or 54 Mbps (802.11b) data rates, or in products containing both bands have.

Those of ordinary skill in the art will understand that information and signals may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that may be referenced in the above description may include voltages, currents, electromagnetic waves, magnetic fields or particles, Particles or particles, or any combination thereof.

Those skilled in the art will appreciate that the various illustrative logical blocks, modules, processors, means, circuits, and algorithm steps described in connection with the embodiments disclosed herein may be implemented or performed with a specific purpose, (Which may be referred to herein as "software") or a combination of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the design constraints imposed on the particular application and the overall system. Those skilled in the art may implement the described functions in various ways for each particular application, but such implementation decisions should not be interpreted as being outside the scope of the present disclosure.

The various embodiments presented herein may be implemented as a method, apparatus, or article of manufacture using standard programming and / or engineering techniques. The term "article of manufacture" includes a computer program, carrier, or media accessible from any computer-readable device. For example, the computer-readable medium can be a magnetic storage device (e.g., a hard disk, a floppy disk, a magnetic strip, etc.), an optical disk (e.g., CD, DVD, etc.), a smart card, But are not limited to, devices (e. G., EEPROM, cards, sticks, key drives, etc.). The various storage media presented herein also include one or more devices and / or other machine-readable media for storing information.

It will be appreciated that the particular order or hierarchy of steps in the presented processes is an example of exemplary approaches. It will be appreciated that, based on design priorities, a particular order or hierarchy of steps in the processes may be rearranged within the scope of this disclosure. The appended method claims provide elements of the various steps in a sample order, but are not meant to be limited to the specific order or hierarchy presented.

The description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present disclosure should not be construed as limited to the embodiments set forth herein, but is to be accorded the widest scope consistent with the principles and novel features presented herein.

Claims (1)

A computer program stored on a computer readable medium.

Images