WO2019039758A1

WO2019039758A1 - Method for generating and learning improved neural network

Info

Publication number: WO2019039758A1
Application number: PCT/KR2018/008621
Authority: WO
Inventors: 김광민
Original assignee: 주식회사 수아랩
Priority date: 2017-08-25
Filing date: 2018-07-30
Publication date: 2019-02-28

Abstract

According to one exemplary embodiment of the present disclosure, a method for generating a neural network model implemented by a computing device may be provided, the method comprising the steps of: generating a first input layer to which a first input is applied; and generating a hidden layer to which an output of the first input layer is applied, wherein the hidden layer includes a second input layer to which a second input is applied. According to one exemplary embodiment of the present disclosure, a method for learning a neural network model implemented by a computing device may be provided, the method comprising the steps of: receiving a training dataset including a first input and a second input; applying the first input to a first input layer; and applying the second input to a second input layer, wherein the second input layer is included in a hidden layer to which an output of the first input layer is applied.

Description

A method for creating and learning an improved neural network

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network, and more particularly, to a neural network in which learning efficiency is improved by processing two or more input data simultaneously.

Deep running, which has been attracting worldwide attention for many years now, has become the most important methodology in artificial intelligence and machine learning. Deep learning is an algorithm for extracting and utilizing features from learning data using deep neural networks (DNN). Each layer of the neural network receives the information extracted from the lower layer and generates more abstract information using this information to transmit to the upper layer. Multiple layers of neural networks can extract very high-level features through multiple abstractions. High-level features include more information than low-level features, but are also robust to variations, so high and stable performance can be achieved if recognition is performed using them.

In the past, it was difficult to use deep neural networks due to various problems such as theoretical problems such as diminishing gradient problem, difficulty in collecting learning data, and limitation of computational performance. However, as new learning algorithms have been developed and techniques for processing large amounts of data have been developed, the theoretical and practical problems have been overcome, and deep running has shown excellent performance in various fields.

In particular, the convolutional neural network (CNN) has become the most important methodology in the field of visual perception, demonstrating the ability to overcome other methodologies in various image recognition contests.

The origins of the convolution neural network are the Neocognitron, which Fukushima imitated in the 1980s. Then, in the 1990s, LeCun began to be widely used in real-world problems by successfully applying a gradient-based learning algorithm. The excellent performance of convolutional neural networks has attracted the explosive interest of many researchers, and research is actively being conducted to further improve the convolution neural network or to apply it to new problems.

Concerning learning, a convoluted neural network requires a large amount of learning data and computational complexity for learning, while humans can learn instantaneously even with only a few examples. To overcome this, a more efficient network structure and learning algorithm are required than nowadays.

The present invention provides a method for generating an improved neural network that enhances learning efficiency by simultaneously processing two or more inputs. The present invention also provides a method for simultaneously learning two or more kinds of inputs to learn a neural network efficiently.

According to an exemplary embodiment of the present disclosure, a method of generating a neural network model implemented by a computing device comprises generating a first input layer to which a first input is applied, step; And generating a hidden layer to which an output of the first input layer is applied, wherein the hidden layer includes a second input layer to which a second input is applied; The method of generating a neural network model including the neural network model may be provided.

In accordance with an exemplary embodiment of the present disclosure, a method for learning a neural network model implemented by a computing device comprises the steps of: (a) generating a training data set comprising a first input and a second input, receiving a training dataset; Applying the first input to a first input layer; And applying the second input to a second input layer, the second input layer being included in a hidden layer to which the output of the first input layer is applied; And a method for learning a neural network model.

According to another exemplary embodiment of the present disclosure, there is provided a computing device comprising: at least one processor; And a memory for storing instructions executable on the one or more processors; The one or more processors generating a first input layer to which a first input is applied; And generating a hidden layer to which the output of the first input layer is applied, the hidden layer including a second input layer to which a second input is applied; A computing device can be provided.

A computer program stored in a computer-readable storage medium including encoded instructions, in accordance with another exemplary embodiment of the present disclosure, when executed by one or more processors of a computer system, Cause the processors to perform operations to create a neural network model, the operations comprising: generating a first input layer to which a first input is applied; And generating a hidden layer to which the output of the first input layer is applied, the hidden layer including a second input layer to which a second input is applied; A computer program stored on a computer-readable storage medium

By this disclosure, two or more types of inputs can be simultaneously processed to create an improved neural network with enhanced learning efficiency. Also, according to the present disclosure, two or more kinds of inputs can be simultaneously processed, and the neural network can be efficiently learned.

1 is a view for explaining the object of the present invention.

2 shows a structure of a neural network model according to an embodiment of the present invention.

3 is a diagram for explaining a method of generating a neural network model according to an embodiment of the present invention.

4 is a diagram for explaining a method for learning a neural network model according to an embodiment of the present invention.

5 illustrates an exemplary block diagram of a neural network module in accordance with an embodiment of the present invention.

6 shows a block diagram of an exemplary computing device for implementing a method for generating a neural network model and a method for learning a neural network, in accordance with an embodiment of the present invention.

Various embodiments and / or aspects are now described with reference to the drawings. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more aspects. However, it will also be appreciated by those of ordinary skill in the art that such aspect (s) may be practiced without these specific details. The following description and the annexed drawings set forth in detail certain illustrative aspects of one or more aspects. It is to be understood, however, that such aspects are illustrative and that some of the various ways of practicing various aspects of the principles of various aspects may be utilized, and that the description set forth is intended to include all such aspects and their equivalents.

In addition, various aspects and features will be presented by a system that may include multiple devices, components and / or modules, and so forth. It should be understood that the various systems may include additional devices, components and / or modules, etc., and / or may not include all of the devices, components, modules, etc. discussed in connection with the drawings Must be understood and understood.

As used herein, the terms " an embodiment, " " an embodiment, " " an embodiment, " " an embodiment ", etc. are intended to indicate that any aspect or design described is better or worse than other aspects or designs. . As used herein, the terms 'component,' 'module,' 'system,' 'interface,' and the like generally refer to a computer-related entity and include, for example, hardware, And software.

In addition, the term " or " is intended to mean " exclusive or " That is, it is intended to mean one of the natural inclusive substitutions " X uses A or B ", unless otherwise specified or unclear in context. That is, X uses A; X uses B; Or when X uses both A and B, " X uses A or B " can be applied to either of these cases. It should also be understood that the term " and / or " as used herein refers to and includes all possible combinations of one or more of the listed related items.

It is also to be understood that the term " comprises " and / or " comprising " means that the feature and / or component is present, but does not exclude the presence or addition of one or more other features, components and / It should be understood that it does not. Also, unless the context clearly dictates otherwise or to the contrary, the singular forms in this specification and claims should generally be construed to mean " one or more. &Quot;

Before describing the embodiments of the present invention in detail, it is to be understood that the present invention is not limited to the above-described embodiments, but may be modified and changed without departing from the scope and spirit of the invention. It is also to be understood that the terminology or words used in the present specification and claims should be interpreted with reference to the meaning of the inventive concept of the present invention based on the principle that the inventor can define the concept of appropriate terms to describe his invention in the best way It should be interpreted as a concept.

1 is a view for explaining the object of the present invention.

A neural network may be a recognition model that mimics the computational capabilities of a biological system using a large number of artificial neurons connected by a link. The neural network may be implemented in software, hardware, or a combination thereof. A neural network may be referred to as an artificial neural network.

Neural networks can utilize artificial neurons that simplify the function of biological neurons. An artificial neuron can be referred to as a node. Artificial neurons can be interconnected via a connection line with connection weights. A connection weight is a specific value of a connection line and may be referred to as a synapse weight or connection strength. Neuronal networks can perform human cognitive or learning processes through artificial neurons.

The neural network can be utilized in the field of determining input data by using a computer. For example, a neural network is used to identify an object existing in a video image, to determine the position of an object, to detect and identify a lesion in a medical image, to understand contents of a voice using voice data, And extracting information for extracting information from the contents of the document. However, the present invention is not limited to this, and the neural network can be utilized in various fields.

The structure of the neural network can be generally determined in advance, and the connection weight according to the connection between the nodes can be calculated to an appropriate value by using data already known to be correct to which class belongs. Data that are known to be correct in this way may be referred to as 'training data', and the process of determining the connection weights may be referred to as 'training'. The neural network for which the connection weight is determined may predict which class the input data belongs to and output the predicted value, which may be referred to as a " test " process. However, the above-described processes can be referred to as various terms.

It may be useful for the neural network to learn additional data of a different type from the learning data. More specifically, the neural network can learn additional features that are difficult to learn using only one type of learning data, using additional data. For example, it may be useful for a neural network to learn discrete data associated with learning data. Here, the discrete data may include data useful for identifying abnormal data among the learning data. In one embodiment, when the neural network is used to check whether defects have occurred in the steel in the steel manufacturing process, the neural network can learn a plurality of steel images 100 that are determined to be defective as learning data . The steel image 100 may be divided into a plurality of image patches 110 and used for learning to reduce the amount of calculation required for learning. For example, in the case where the steel image 100 is a 1000 x 200 pixel size, the steel image 100 may be divided into a plurality of image patches 110 having a pixel size of 20x20 and used for learning. In this case, the plurality of image patches 110 may have respective coordinate values. For example, as shown in Fig. 1, defects of steel may exist in an image patch whose coordinate values are (7, 4), (7, 5). In the process of manufacturing steel, if defects occur at a certain position (for example, the coordinate values of image patches are (7, 4), (7, 5)), The coordinate value of the image patch 110 may be provided as learning data so that it can learn which coordinate value the image patch (i.e., abnormal data) has. This can increase the learning efficiency of the neural network.

For this purpose, the neural network model of the present invention may include a first input layer 220 to which a first input is applied and a second input layer 270 to which a second input is applied.

The first input 200 applied to the first input layer 220 may include general learning data. More specifically, the first input 200 may include learning data that has been used in learning of a conventional neural network. For example, the first input 200 may include image data, sound data, image data, character data, and the like. However, the present invention is not limited to this, and the first input may include various data.

The second input 210 applied to the second input layer 270 may include a different type of data than the first input 200. Specifically, the second input may include data capable of learning characteristics that are difficult to learn with only the first input 200. [ For example, the second input may be applied with discrete data related to the first input. In one embodiment, when a plurality of image patches 110 are applied to the first input layer 220 as a first input 200, 2 input 210, the coordinate value of the image patch 110 may be applied. When the coordinate values of the image patches in which defects mainly occur in the plurality of image patches 110 included in the steel image 100 are (7, 4) and (7, 5) (7, 4) and (7, 5) by using the coordinate values of the image patch 110 as the second input together with the image patch 110 of the image patch 110 as the learning data, You can learn. Neural networks can not learn these features with only the first input.

The technical features of the present invention can be used to improve learning efficiency by providing more data to the neural network. The neural network model of the present invention has a structure for simultaneously processing two kinds of data which are different from each other, so that efficient learning is possible.

In one embodiment of the invention, the neural network model may comprise a plurality of layers. For example, the neural network model may include a first input layer 220, a hidden layer 230, and an output layer (not shown). The first input layer 220 may receive a first input for performing learning and transmit the first input to the hidden layer 230. The output layer may transmit the first input to the hidden layer 230 based on data received from the hidden layer 230, The output of the model can be generated. The hidden layer 230 is located between the first input layer 220 and the output layer and can change learning data transmitted through the first input layer 220 to a value that is easy to predict.

The layers included in the first input layer 220, the hidden layer 230, and the output layer may include a plurality of nodes. A node included in the first input layer 220 is referred to as a first input node, a node included in the hidden layer 230 is referred to as a hidden node, A node may be referred to as an output node.

The first input nodes included in the first input layer 220 and the hidden nodes included in the hidden layer 230 may be connected to each other through connection links having connection weights. The hidden nodes included in the hidden layer 230 and the output nodes included in the output layer may be connected to each other through connection lines having connection weights.

The hidden layer 230 may include a plurality of layers. A neural network in which the hidden layer 230 includes a plurality of layers may be referred to as a deep neural network. Learning a deep neural network can be referred to as deep learning.

The hidden layer 230 may include a plurality of convolution layers 240, a pooling layer 250, and a fully connected

layer

260 and 270. Convolution layer 240 may refer to a layer having a structure in which neighboring nodes in the previous layer are connected to nodes in the next layer. The pooling layer 250 may refer to a layer having a structure capable of selecting and integrating a part at a node of a previous layer. The pulley connected

layers

260 and 280 may refer to a layer having a structure in which the nodes included in the layer are connected to all nodes included in the previous layer. However, the above-mentioned layers are exemplary, and some of the layers may be changed or omitted, and additional layers may be included in the hidden layer 230. In addition, some layers included in the hidden layer 230 may be included in the first input layer 220 or the output layer as the case may be.

In an embodiment of the present invention, the hidden layer 230 may include a second input layer 270. The second input layer 270 may be disposed in parallel with any one of the pulley connected layers included in the hidden layer 230. Also, the second input layer 270 may be applied with the second input 210, and may transmit the output to the upper layer.

In one embodiment of the present invention, the neural network model of the present invention may include a first input layer 220, which is a first input, and a second input layer 270, to which a second input is applied.

The second input 210 applied to the second input layer 270 may include a different type of data than the first input 200. More specifically, the second input may include data that can learn features that are difficult to learn with only the first input 200. For example, the second input may include data that is applied to discrete data associated with the first input . In one embodiment, when a plurality of image patches 110 are applied to the first input layer 220 as a first input 200, 2 input 210, the coordinate value of the image patch 110 may be applied. When the coordinate values of the image patches in which defects mainly occur in the plurality of image patches 110 included in the steel image 100 are (7, 4) and (7, 5) (7, 4) and (7, 5) by using the coordinate values of the image patch 110 as the second input together with the image patch 110 of the image patch 110 as the learning data, You can learn. Neural networks are difficult to learn these characteristics with only the first input

In one embodiment of the present invention, the neural network model may include a first input layer 220, a hidden layer 230, and an output layer. However, the above-mentioned layers are exemplary, and some of the layers may be changed or omitted, and additional layers may be included in the neural network model. The layers may refer to multiple layers.

In one embodiment of the invention, the neural network model may be a convolution neural network. A convolutional neural network is a feedforward artificial neural network (a network in which data is handed over from the lower layer to the upper layer) tied in a manner that responds to overlapping regions of individual neurons, Lt; / RTI >

In one embodiment of the present invention, a method of generating a neural network model may include generating (300) a first input layer 220 to which a first input 200 is applied.

The first input layer 220 may be generated as the lowest layer of the neural network model. The first input layer 220 may receive the first input 200. The first input layer 220 may communicate the received first input 200 to the hidden layer 230 for use in learning and testing. The first input 200 may be transformed to be processed in the hidden layer 230 in the first input layer 220. The first input 200 may also be preprocessed to improve learning efficiency. For example, the first input 200 may be preprocessed such that the noise component is removed or upscaled. However, the present invention is not limited to this, and the first input layer 220 may be formed as a lower layer or a higher layer of the neural network model, and may perform various operations.

In one embodiment of the present invention, a method of generating a neural network model may include generating 310 a hidden layer 230 to which the output of the first input layer 220 is applied. The hidden layer 230 may include a second input layer 270 to which a second input 210 is applied.

The hidden layer 230 may include a plurality of layers. For example, the hidden layer 230 may include at least one convolution layer 240, at least one pulling layer 250, at least one pulley connected

layer

260, 280, and a second input layer 270 . The plurality of layers included in the hidden layer 230 may be arranged in series or in parallel with each other. However, the present invention is not limited thereto, and the hidden layer 230 may include various layers.

The hidden layer 230 may receive the output of the first input layer 220 and extract and classify the features. More specifically, the convolution layer 240 can extract features from the hidden layer 230 by abstracting the input data. For example, the convolution layer 240 may apply a filter to a certain area of data (e.g., 5x5 or 3x3) to extract features. These filters can be changed according to learning. The pulling layer 250 may perform a sub-sampling process. The sub-sampling process can take a maximum value or take an average value for the input data. The neural network model may obtain a global feature while passing the first input 200 through the multiple layers of convolution layer 240 and pooling layer 250. However, the present invention is not limited to this, and the hidden layer can perform various operations.

The pulley connected

layers

260 and 280 are generally positioned above the convolution layer 240 and the pulling layer 250 and perform operations of sorting using extracted features while passing through the convolution layer and the pooling layer . The pulley connected

layers

260 and 280 may be formed in multiple layers. For example, the pulley connected

layers

260 and 280 may be formed in two layers as shown in FIG. However, the present invention is not limited to this, and the pulley-connected

layers

260 and 280 may be formed in three or more layers.

The second input layer 270 may receive the second input 210. The second input layer 270 may be used to classify the received second input 210 with the extracted features as the first input 200 passes through the convolution layer 240 and the pulling layer 250 can do. To this end, the second input layer 270 may be created above the convolution layer 240 and the pooling layer 250. However, the second input layer 270 may be formed at various positions on the hidden layer 230.

In an embodiment of the invention, the second input layer may receive the second input directly. For example, the second input layer can receive the second input directly without going through another layer. However, the present invention is not limited to this, and the second input layer can receive the second input through another layer. For example, the second input layer may receive its output as an input after the second input is applied to another layer. In this case, the node to which the second input is applied can be connected to only a small number of nodes included in the upper layer. For example, a node to which a second input is applied may be connected to only one node included in the upper layer. However, the present invention is not limited to this, and a node to which a second input is applied may be connected to two or more nodes included in an upper layer. Also, the node through which the second input is passed may not be connected to the node used to extract the characteristics of the first input. Therefore, the second input may be rarely used in the process of extracting the characteristic of the first input, or may be used only in a part of the process.

In an embodiment of the present invention, the second input layer 270 may be disposed in parallel with the first pulley-connected layer 260 included in the pulley-connected

layers

260 and 280.

The neural network may make the second input applied to the second input layer 270 available for classification operations. More specifically, the second input layer 270 may include a convolution layer 240 and an upper layer of the pulling layer 250 such that the second input 210 is not used in the process of extracting the characteristics of the first input 200. [ May be disposed in parallel with a first pulley-connected layer located in the first pulley. As described above, the pulley connected

layers

260 and 280 are generally positioned above the convolution layer 240 and the pulling layer 250 to perform the task of classifying extracted features. Thus, the second input layer 270 may be disposed in parallel with the first pulley connected layer 260 to allow the second input 210 to be used to perform the classification operation. The second input layer may be disposed in parallel with any of the convolution layer 240 and the pulling layer 250 and the second input 210 may have a characteristic of the first input 200 Can be used in the extraction process.

In an embodiment of the present invention, the first pulley connected layer 260 may be the lowest layer among the pulley connected

layers

260 and 280.

As described above, it may be effective that the second input 210 is applied after the feature of the first input 200 is extracted rather than in the process of extracting the feature of the first input 200. [ Also, it may be efficient to apply the second input 210 to the step of starting the task of classifying using the extracted features. The pulley connected

layers

260 and 280 may be formed in multiple layers for efficient classification generally. Accordingly, the second input layer 270 is arranged in parallel to the lowermost layer among the plurality of pulley-connected layers, and is used for starting the operation of classifying the second input 210 using the extracted feature . However, the present invention is not limited thereto. In some cases, the second input layer 270 may be disposed in a lower layer or a higher layer among the pulley-connected layers in parallel.

In an embodiment of the present invention, the second input layer 270 and the first pulley connected layer may be connected to the second pulley connected layer. Where the second pulley connected layer may be a layer located on top of the second input layer 270 and the first pulley connected layer.

For efficient learning and testing, the second input 210 may be applied to multiple layers of pulley-connected layers with features extracted from the lower layer and used for classification operations. Thus, the second input layer 270 and the first pulley connected layer can be coupled to deliver the extracted features of the first input 200 and the second input 210 to the second pulley connected layer. The second pulley connected layer can then process the received data and pass it back to the multi-layer pulley connected or output layer. The second input layer 270 and the first pulley connected layer 260 may be connected to various kinds of layers. For example, the second input layer 270 and the first pulley connected layer may be coupled to a 1x1 convolution layer. Here, the 1x1 convolution layer can perform the task of classifying extracted features, such as pulley connected

layers

260 and 280, but is not limited thereto.

In one embodiment of the present invention, the first input 200 and the second input 210 may comprise different types of data.

As described above, the first input 200 applied to the first input layer 220 may include general learning data. More specifically, the first input 200 may include learning data that has been used in learning of a conventional neural network. For example, the first input 200 may include image data, sound data, image data, character data, and the like. However, the present invention is not limited to this, and the first input may include various data.

The second input 210 applied to the second input layer 270 may include a different kind of data than the first input 200. [ Specifically, the second input may include data capable of learning characteristics that are difficult to learn with only the first input 200. [ For example, the second input may be applied with discrete data related to the first input. In one embodiment, when a plurality of image patches 110 are applied to the first input layer 220 as a first input 200, 2 input 210, the coordinate value of the image patch 110 may be applied. When the coordinate values of the image patches in which defects mainly occur in the plurality of image patches 110 included in the steel image 100 are (7, 4) and (7, 5) (7, 4) and (7, 5) by using the coordinate values of the image patch 110 as the second input together with the image patch 110 of the image patch 110 as the learning data, You can learn. Neural networks are difficult to learn these characteristics with only the first input

In one embodiment of the present invention, the second input 210 may include discrete data related to the first input 200. More specifically, the discrete data associated with the first input 200 can be used to learn characteristics that are difficult to learn with only the first input. In one embodiment, the discrete data may include data useful for identifying normal and abnormal data in the training data applied to the first input 200. [ Specifically, in the case where the first input 200 is a plurality of image patches 110 in which the steel image 100 is divided, the discrete data related to the first input 200 may include defects (I.e., normal data), and data related to a cause of occurrence of abnormal data that can be used to identify an image patch in which a defect exists (i.e., abnormal data). For example, discrete data may include coordinate values of image patches or the like that can be used to identify locations where defects predominantly occur in steel, or steel images that may be used to identify defects in certain processes May include data such as the time, place, process information, etc. at which the apparatus 100 is photographed, and may include data such as temperature, humidity, manufacturing apparatus, etc., which can be used to identify defects in any manufacturing environment.

In another embodiment, the discrete data may include data indicative of the characteristics of the training data applied to the first input 200. More specifically, when the first input 200 is speech data, the discrete data related to the first input 200 may include the sex, age, key (s) of the speaker that can be used to learn the characteristics of the speech data according to the speaker , Weight, and the like, and may include a recording location, a recording time, a recording device, etc., which can be used to learn characteristics of voice data according to the surrounding environment. The discrete data associated with the first input 200 may include various data.

The second input 210, which includes discrete data related to the first input 200, can be usefully used to learn the first input 200. In particular, the second input may be used as additional information to learn the first input 200 and may be used to learn additional features of the first input 200 associated with the second input. For example, if the first input 200 is the partitioned image patch 110 of the steel image 100 and the second input 210 is the coordinate value of the image patch 110, The position where the defects mainly occur in the steel image 100 can be learned by using the coordinate values of the defects 110. As another example, if the first input 200 is speech data and the second input 210 is the speaker's gender and age, the neural network model can learn the characteristics of speech data according to the speaker's gender and age. The neural network model of the present invention has a structure for simultaneously processing two kinds of data which are different from each other, thereby enabling efficient learning.

In one embodiment of the present invention, a method for learning a neural network model includes receiving (400) a learning data set that includes a first input (200) and a second input (210), a first input layer (410) applying a first input (200) to the second input layer (270) and applying (420) a second input to a second input layer (270). Here, the second input layer 270 may be included in the hidden layer 230 to which the output of the first input layer 220 is applied.

The neural network model of the present invention can be learned by various methods. For example, neural network models can be learned through supervised learning. The supervised learning is a technique of inputting the learning data and the corresponding output data together into the convolutional neural network and updating the weight so that the output data corresponding to the learning data is output. Specifically, the neural network model of the present invention can update connection weights through back-propagation learning. The error-domain propagation learning estimates an error through forward computation on a given learning data, and then estimates an error in a reverse direction starting from the output layer and toward the first input layer 220 and the second input layer 270 It can be a way to update the connection weights in the direction of reducing the error while propagating one error. However, the present invention is not limited to this, and the neural network can be learned through Unsupervised Learning or Reinforcement Learning, or may be learned by other learning methods.

The learning data set of the present invention may include a first input 200 and a second input 210. The first input 200 can be applied to the first input layer 220 and the second input 210 can be applied to the second input layer 270 included in the hidden layer 230 . In one embodiment, the second input layer 270 may be disposed in parallel with the first pulley connected layer included in the pulley connected layer. Here, the first pulley-connected layer may be the lowest layer among the pulley-connected layers. The second input layer 270 and the first pulley connected layer may be connected to a second pulley connected layer, wherein the second pulley connected layer includes a second input layer 270 and the first pulley connected layer The layer may be a layer located at the top of the layer.

As described above, the first input 200 may include general types of data used as learning data of a neural network. For example, the first input 200 may include image data, sound data, image data, character data, and the like. The first input 200 may include various types of data that can be used for learning of the neural network model.

The second input may include a different type of data than the first input 200 applied to the first input layer 220. For example, if the first input 200 includes image data, the second input 210 may include other types of data such as sound data, image data, character data, and the like. The second input 210 may include data of the same kind as that of the first input 200, and may also include various other types of data.

The second input 210 may include discrete data related to the first input 200. The discrete data associated with the first input 200 may be data including information regarding the first input 200. [ In one embodiment, in the case where the first input 200 is a plurality of image patches 110 that have segmented the steel image 100, the discrete data associated with the first input 200 is the coordinate value of the image patch, The time, place and temperature at which the camera 100 is photographed, process information, and the like. In another embodiment, when the first input 200 is voice data, the discrete data associated with the first input 200 may be the gender, age, recording location, recording time, recording device, etc. of the speaker.

The second input 210, which includes discrete data related to the first input 200, can be usefully used to learn and test the first input 200. For example, if the first input 200 is the partitioned image patch 110 of the steel image 100 and the second input 210 is the coordinate value of the image patch 110, The position where the defects mainly occur in the steel image 100 can be learned by using the coordinate values of the defects 110. As another example, if the first input 200 is speech data and the second input 210 is the speaker's gender and age, the neural network model can learn the characteristics of speech data according to the speaker's gender and age. The neural network model of the present invention has a structure for simultaneously processing two kinds of data which are different from each other, thereby enabling efficient learning.

5 shows an exemplary block diagram of a neural network module in accordance with an embodiment of the present invention.

In an embodiment of the present invention, the neural network module 600 may include a neural network generator 610 and a neural network learner 620. The neural network generating unit 610 may include a first input layer generating unit 611 and a hidden layer generating unit 612. The hidden layer generating unit 612 may include a second input layer generating unit 613. The components of neural network module 600 shown in FIG. 5 are exemplary, some of which may be omitted, or additional components may be present.

In an embodiment of the present invention, the first input layer generator 611 may generate the first input layer 220 to which the first input 200 is applied. The hidden layer generating unit 612 may generate the hidden layer 230 by generating the convolution layer 240, the pulling layer 250, the pulley connected

layers

260 and 280, and other layers. The second input layer generating unit 613 may generate the second input layer 270. [

In one embodiment of the present invention, the neural network learning unit 620 may receive a learning data set that includes a first input 200 and a second input 210. [ The neural network learning unit 620 applies a first input 200 to the first input layer 220 and a second input 210 to the second input layer 270 in order to learn a neural network model. can do.

Although the present invention has been described above generally in terms of features that may be executed on a computer or processor within one or more servers, those skilled in the art will appreciate that the invention may be implemented in combination with other program modules and / You will know.

Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. It will also be appreciated by those skilled in the art that the methods of the present invention may be practiced with other computer systems, such as 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 invention 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 nonvolatile 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. The computer-readable storage medium may be any form of storage device such as RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital video disk (DVD) or other optical disk storage, magnetic cassette, magnetic tape, Apparatus, or any other medium that can be used to store the desired information that can be accessed by a computer.

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, the transmit and receive (communication) media include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared and 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.

6, there is shown an exemplary environment for implementing various aspects of the invention including a computing device 702 for implementing a neural network model of the present invention and a method for learning a neural network, Apparatus 702 includes a processing unit 704, a system memory 706, and a system bus 708. The system bus 708 couples system components, including but not limited to, system memory 706 to the processing unit 704. The processing unit 704 may be any of a variety of commercially available processors. Dual processors and other multiprocessor architectures may also be used as the processing unit 704.

The system bus 708 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 706 includes a read only memory (ROM) 710 and a random access memory (RAM) 712. The basic input / output system (BIOS) is stored in a non-volatile memory 710, such as a ROM, EPROM, or EEPROM, which can be used to assist in transferring information between components within the computing device 702, It contains basic routines. The RAM 712 may also include a high speed RAM such as static RAM for caching data.

The computing device 702 may also be configured for external use within an appropriate chassis (not shown), such as an internal hard disk drive (HDD) 714 (e.g., EIDE, SATA) (E.g., read from or write to a removable diskette 718), and an optical disk drive 720 (e.g., a CD-ROM For reading disc 722 or reading from or writing to other high capacity optical media such as DVD). The hard disk drive 714, magnetic disk drive 716 and optical disk drive 720 are connected to the system bus 708 by a hard disk drive interface 724, a magnetic disk drive interface 726 and an optical drive interface 728, respectively. . The interface 724 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 storage media provide nonvolatile storage of data, data structures, computer-executable instructions, and the like. In the case of computing device 702, the drives and media correspond to storing any data in a suitable digital format. Although the description of the above computer-readable storage medium refers to a removable optical medium such as a HDD, a removable magnetic disk, and a CD or a DVD, those skilled in the art will appreciate that a zip drive, magnetic cassette, flash memory card, 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 invention.

A number of program modules, including an operating system 730, one or more application programs 732, other program modules 734, and program data 736, may be stored in the drive 732 and the RAM 712. Here, the program module 734 may include a neural network module (not shown) for generating, learning, and testing a neural network model. All or a portion of the operating system, applications, modules, and / or data may also be cached in RAM 712. It will be appreciated that the present invention may be implemented in a variety of commercially available operating systems or combinations of operating systems.

A user may enter commands and information into computing device 702 via one or more wired / wireless input devices, e.g., a pointing device such as keyboard 738 and mouse 740. [ Other input devices (not shown) may include a microphone, IR remote control, joystick, game pad, stylus pen, touch screen, etc. These and other input devices are often connected to the processing unit 704 via an input device interface 742 that is coupled to the system bus 708, but may be a parallel port, an IEEE 1394 serial port, a game port, a USB port, an IR interface, &Lt; / RTI > and so forth.

A monitor 744 or other type of display device is also connected to the system bus 708 via an interface, such as a video adapter 746. [ In addition to the monitor 744, the computer typically includes other peripheral output devices (not shown) such as speakers, printers,

Computing device 702 may operate in a networked environment using logical connections to one or more remote computers, such as remote computer (s) 748 via wired and / or wireless communication. The remote computer (s) 748 may be a workstation, a server 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 750 is shown. The logical connections depicted include a wired / wireless connection to a local area network (LAN) 752 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 computing device 702 is connected to the local network 752 via a wired and / or wireless communication network interface or adapter 756. The adapter 756 may facilitate wired or wireless communication to the LAN 752 and the LAN 752 also includes a wireless access point installed therein to communicate with the wireless adapter 756. When used in a WAN networking environment, the computing device 702 may include a modem 758, or may be connected to a communications server on the WAN 754, or to establish communications over the WAN 754, And other means. A modem 758, which may be an internal or external and a wired or wireless device, is connected to the system bus 708 via a serial port interface 742. In a networked environment, program modules depicted relative to computing device 702, or portions thereof, may be stored in remote memory / storage device 750. 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 computing device 702 may be any wireless device or entity that is deployed and operated in wireless communication, such as a printer, a scanner, a desktop and / or portable computer, a portable data assistant (PDA) To any associated equipment or location, and to communication with the phone. 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 802.6 (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 can operate in unlicensed 2.4 and 5 GHz wireless bands, for example, at 6 Mbps (802.6a) or 54 Mbps (802.6b) data rates, or in products containing both bands have.

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 embodied directly in electronic hardware, (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 functionality in various ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

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 term " machine-readable medium " includes, but is not limited to, various other mediums capable of storing and / or retaining instruction (s) and / or data.

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, certain orders or hierarchies of steps in processes may be rearranged within the scope of the present invention. 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 present invention. 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 invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features presented herein.

As described above, relevant contents have been described in the best mode for carrying out the invention.

The present invention can be used for learning of a neural network utilized in a field for discriminating input data using a computing device.

Claims

A method of generating a neural network model implemented by a computing device,

Generating a first input layer to which a first input is applied; And

Generating a hidden layer to which an output of the first input layer is applied, the hidden layer including a second input layer to which a second input is applied;

Containing

A method for generating a neural network model.
The method according to claim 1,

The hidden layer includes:

At least one fully connected layer;

Lt; / RTI >

The second input layer comprising:

A first Fully-Connected Layer included in the at least one pulley-connected layer;

How to create a neural network model
3. The method of claim 2,

Wherein the first pulley connected layer comprises:

Wherein at least one of the at least one pulley-

A method for generating a neural network model.
3. The method of claim 2,

The second input layer and the first pulley connected layer,

Connected to a second pulley connected layer, the second pulley connected layer being a layer located on top of the second input layer and the first pulley connected layer;

A method for generating a neural network model.
The method according to claim 1,

Wherein the first input and the second input comprise different types of data,

A method for generating a neural network model.
The method according to claim 1,

Wherein the second input comprises discrete data related to the first input,

A method for generating a neural network model.
The method according to claim 1,

The neural network model is a convolutional neural network,

A method for generating a neural network model.
A method for learning a neural network model implemented by a computing device,

Receiving training data comprising a first input and a second input;

Applying the first input to a first input layer; And

Applying the second input to a second input layer, the second input layer being included in a hidden layer to which the output of the first input layer is applied;

/ RTI >

A method for learning neural network model.
13. A computing device,

One or more processors; And

A memory for storing instructions executable on the one or more processors;

Lt; / RTI >

Wherein the one or more processors comprise:

Generating a first input layer to which a first input is applied; And

Generating a hidden layer to which an output of the first input layer is applied, the hidden layer including a second input layer to which a second input is applied;

Computing device.
A computer program stored in a computer-readable storage medium including encoded instructions, the computer program being executable by one or more processors of a computer system to cause the one or more processors to generate a neural network model , Said operations comprising:

Generating a first input layer to which a first input is applied; And

Generating a hidden layer to which an output of the first input layer is applied, the hidden layer including a second input layer to which a second input is applied;

Containing

A computer program stored in a computer-readable storage medium.