KR20200023695A - Learning system to reduce computation volume - Google Patents

Abstract

The present invention relates to a learning system for reducing an operation amount, and more particularly, to an operation amount reducing system that creates a learning model which realizes optimal performance while reducing an excessive operation amount when performing 3D-based deep learning. The operation amount reducing system includes a plurality of cells including a plurality of blocks, and fixes connection relationships of input and output between cells to maintain a specific overall system structure. The block includes a plurality of layers that perform data input, processing or output, and the plurality of blocks or layers can change structures thereof during a data learning process.

Description

Learning system to reduce the amount of computation {LEARNING SYSTEM TO REDUCE COMPUTATION VOLUME}

The present invention relates to a learning system for reducing the amount of computation, and more particularly, to a computational amount reduction system for creating a learning model that achieves optimal performance while reducing excessively large amounts of computation in three-dimensional deep learning.

Neural networks represent data structures that model the living brain. In neural networks, nodes may represent neurons that are interconnected and operate collectively to process input data. Examples of different types of neural networks include, but are not limited to, convolutional neural networks, recurrent neural networks, deep belief networks, and restricted boltzamann machines. This is not restrictive.

Neural networks can be used to extract "features" from complex input data. The neural network may include a plurality of layers. Each layer may receive input data and process the received input data to generate output data. The output data may be a feature map of the input data. The neural network may be used to generate a feature map of the input data by convolving the input image or feature map with the convolution kernel. The initial layer of the neural network may operate to extract low level features such as edges and / or gradients from inputs such as images. Subsequent layers of the neural network may extract progressively more complex features, such as the eyes or the nose.

Korean Laid-Open Patent Publication No. 10-2017-0083419, 2017.07.18

The problem to be solved by the present invention is to provide a system for reducing the amount of calculation and increase the accuracy in the calculation of the three-dimensional learning model.

In addition, the problem to be solved by the present invention is to provide a system for enhancing the accuracy of the region segmentation to be divided in the three-dimensional image segmentation.

Problems to be solved by the present invention are not limited to the above-mentioned problems, and other problems not mentioned will be clearly understood by those skilled in the art from the following description.

Learning system for reducing the amount of calculation according to an embodiment of the present invention for solving the above problems includes a plurality of cells including a plurality of blocks, maintaining the overall system structure specifying the connection relationship between the input and output between each cell The block may include a plurality of layers for inputting, processing, or outputting data, and the plurality of blocks or layers may be structurally changed in a data learning process.

In the case where the nth output is generated by the first cell among the plurality of cells, an input capable of entering the convolutional layer inside the block may include an n-1th cell output, an n-2th cell output, or the like. It is characterized in that the output from the block belonging to the first cell.

In the learning system for reducing the amount of computation according to an embodiment of the present invention, after only one cell of the plurality of cells has evolved, the number of blocks of all other cells is equally applied to the number of blocks of one cell. It is done.

The learning system for reducing the amount of computation according to an embodiment of the present invention is characterized in that the amount of computation is reduced through network pruning for a plurality of convolutional layers among the plurality of layers.

The network pruning of the plurality of convolutional layers includes deleting all connections having a weight less than or equal to a predetermined value for each of the plurality of convolutional layers, and zero input weights for each of the plurality of convolutional layers. Or all of the one or more nodes having zero output weights are deleted, and the nodes of the n-1 th layer and the n + 1 th layer of the same position where the node deleted in the n th layer among the plurality of convolutional layers are located are connected. It is characterized in that the connection to send the output of the n-1 th layer to the n + 1 th layer to be added.

The deletion of all the connections having a weight less than or equal to the predetermined value is performed for each of the plurality of layers, and the deletion of each layer connection is sequentially performed. The weight less than or equal to the predetermined value in the first layer is performed. If all connections with a value are deleted, the first layer from which the connection is deleted is re-learned, and if the performance of the network is changed by the re-learning, all the connections with a weight less than or equal to a predetermined value in the second layer are deleted. It is characterized by.

The learning system for reducing the amount of computation according to an embodiment of the present invention is characterized in that the accuracy of image segmentation is increased by modifying and calculating the loss function combined with the loss function for the plurality of layers.

The loss function is modified to be calculated as a combined loss function, and the similarity between the first region and the second region is calculated as a mixed matrix, and a dice value is measured, and the measured dice values are mixed to combine the loss function. And modified as a loss function and class balanced through weighted cross entropy, wherein the first region is an area extracted from the original image, the second region is an area extracted from the label image, and the first area and the The second area is an area of the same size in each of the original image and the label image.

The combined loss function satisfies Equation 1 below.

<Equation 1>

In Equation 1, L _{CE + dice} is a loss value reflecting cross entropy and a dice value, dice is a value obtained by calculating a similarity between the first area and the second area by a mixed matrix, and λ _CE is a value for the cross entropy loss function. The weighting parameter, λ _dice, is a weighting parameter for the die loss function, and L _dice is a loss value that reflects the dice value.

Other specific details of the invention are included in the detailed description and drawings.

According to the present invention, in the calculation of the three-dimensional learning model, by using the combined loss function while fixing the connection relationship between the cells so that structural changes do not occur, the optimal performance is quickly derived while using less computational amount. can do.

In addition, according to the present invention, by fixing the connection between the input and output between the cells, the model can be evolved while maintaining the structure of the seed model.

In addition, according to the present invention, in the calculation of the three-dimensional learning model, it is possible to reduce the amount of calculation through the pruning of the network.

In addition, according to the present invention, the loss function is corrected and calculated as a combined loss function, so that the accuracy of region segmentation in an image can be increased.

Effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a diagram illustrating a connection relationship between an input and an output between each cell of a learning system that reduces an amount of computation according to an exemplary embodiment of the present invention.
2 to 4 are diagrams for explaining an example of an input and output relationship in which each cell evolves in a learning system for reducing the amount of computation according to an embodiment of the present invention.
FIG. 5 illustrates a node of an n-1 th layer and n + 1 in the same position where a node deleted from an n th layer is located among a plurality of convolutional layers in a network pruning of a learning system that reduces an amount of computation according to an exemplary embodiment of the present invention. A diagram for describing a connection for sending an output of an n-1 th layer to an n + 1 th layer so that nodes of a first layer are connected.
FIG. 6 is a diagram illustrating a result of correcting a loss function of a learning system that reduces a computation amount according to an embodiment of the present invention and reflecting it on a medical image.
7 is a graph showing the result of the learning system to reduce the amount of calculation according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them will be apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be embodied in various different forms, only the present embodiments are intended to complete the disclosure of the present invention, and those skilled in the art to which the present invention belongs. It is provided to fully inform the skilled person of the scope of the invention, which is defined only by the scope of the claims.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, “comprises” and / or “comprising” does not exclude the presence or addition of one or more other components in addition to the mentioned components. Like reference numerals refer to like elements throughout, and "and / or" includes each and all combinations of one or more of the mentioned components. Although "first", "second", etc. are used to describe various components, these components are of course not limited by these terms. These terms are only used to distinguish one component from another. Therefore, of course, the first component mentioned below may be the second component within the technical spirit of the present invention.

Unless otherwise defined, all terms used in the present specification (including technical and scientific terms) may be used in a sense that can be commonly understood by those skilled in the art. In addition, terms that are defined in a commonly used dictionary are not ideally or excessively interpreted unless they are specifically defined clearly.

A deep neural network (DNN) according to embodiments of the present invention refers to a system or a network that performs a determination based on a plurality of data by constructing one or more layers in one or more computers. For example, the deep neural network may be implemented as a set of layers including a convolutional pooling layer, a locally-connected layer, and a fully-connected layer. The convolution pooling layer or the local access layer can be configured to extract features in the image. The fully connected layer may determine the correlation between the features of the image. In some embodiments, the overall structure of the deep neural network may be in the form of a local connection layer followed by a convolution pooling layer and a fully connected layer in the local connection layer. The deep neural network may include various criteria (ie, parameters), and may add new criteria (ie, parameters) through an input image analysis.

The deep neural network according to the embodiments of the present invention is a structure called a convolutional neural network suitable for image analysis, and is a feature extraction layer for self-learning a feature having the highest discriminative power from given image data. ) And a prediction layer that learns a prediction model to have the highest prediction performance based on the extracted feature.

The feature extraction layer spatially integrates a convolution layer and a feature map to create a feature map by applying a plurality of filters to each region of an image, thereby allowing the feature to be changed to change in position or rotation. It can be formed in a structure that repeats the alternating layer (Pooling Layer) to be extracted several times. Through this, it is possible to extract various levels of features, from low level features such as points, lines and planes to complex high level features.

The convolutional layer obtains a feature map by taking a nonlinear activation function on the inner product of the filter and the local receptive field for each patch of the input image. In comparison, CNN is characterized by the use of filters with sparse connectivity and shared weights. This connection structure reduces the number of parameters to be learned and makes the learning through the backpropagation algorithm more efficient, resulting in better prediction performance.

The pooling layer or sub-sampling layer creates a new feature map using the local information of the feature map obtained from the previous convolutional layer. In general, the feature map newly created by the integration layer is reduced to a smaller size than the original feature map. The typical integration method includes the maximum pooling to select the maximum value of the corresponding area in the feature map and the corresponding feature map in the feature map. Average pooling, which finds the mean of a domain. The feature map of the unified layer is generally less affected by the location of any structure or pattern present in the input image than the feature map of the previous layer. That is, the integrated layer can extract features that are more robust to regional changes such as noise or distortion in the input image or previous feature maps, and these features can play an important role in classification performance. Another role of the unified layer is to allow the deeper structure to reflect the characteristics of a wider area as it goes up to the upper learning layer. As feature extraction layers accumulate, the lower layer reflects local features and rises to the upper layer. Increasingly, it is possible to generate a feature that reflects a feature of the entire abstract image.

As such, the features finally extracted through the repetition of the convolutional layer and the integrated layer are classified into a classification model such as Multi-layer Perception (MLP) or Support Vector Machine (SVM). Combined in the form of -connected layer, it can be used for classification model training and prediction.

However, the structure of the deep neural network according to the embodiments of the present invention is not limited thereto, and may be formed of a neural network having various structures.

Reference numerals in the drawings used in the present specification apply only to the corresponding drawings described, and do not use different drawings and reference numerals.

Hereinafter, a learning system for reducing the amount of calculation according to embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The learning system for reducing the amount of computation according to an embodiment of the present invention includes a plurality of cells including a plurality of blocks, and the plurality of blocks includes a plurality of layers.

The learning system that reduces the amount of computation of the present invention is fixed so as to maintain the structure of the entire system in which the connection relationship between the input and the output between each cell is specified.

The plurality of layers perform input, processing or output of data.

The layers exemplified in the specification of the present invention may include 1X1X1 convolution layer, 3X3X3 convolution layer, 5X5X5 convolution layer, 7X7X7 convolution layer, 3X3X3 max pooling layer, 3X3X3 average consolidation layer, Although the 3X3X3 atrous convolution layer and the 1X1X7 1X7X1 7X1X1 convolution layer are not limited to the above examples, all layers that can be configured and various kinds of layers are included.

1 is a view for explaining the connection between the input and output between each cell of the learning system to reduce the amount of calculation according to an embodiment of the present invention.

Referring to FIG. 1, a cell is configured from the first cell 100 to the sixth cell 600, and each cell includes two layers.

The learning system that reduces the amount of computation of the present invention fixes the input and output connections between cells in order to maintain the structure. When the connection between cells is fixed and learned, no structural change occurs. The structure of each block (not shown) can be changed, and each layer can also be changed.

In the case where the structural change does not occur by fixing the connection relationship between cells as in the present invention, the structure of the seed model may be maintained while the model may evolve.

In an embodiment of the present invention, in FIG. 1, an input value is input to a first cell, a second cell 200 in a first cell 100, a third cell 300 in a second cell 200, and a third cell. The third cell 300 is connected to the block, the fourth cell 400 in the block, the fourth cell 400 in the fifth cell 500, the fifth cell 500 in the sixth cell 600 is connected to the sixth cell At 600 an output value is output.

When the connection between the layers is sequentially described, an input value is input to the first layer 111 of the first cell 100, the second layer 112, the second layer 112, the first layer 111, and the second layer 112, respectively. Third layer 211 of second cell 200, fourth layer 212 at third layer 211, fifth layer 311, third layer of third cell 300 at fourth layer 212 5th layer 311 to 6th layer 312, 6th layer 312 to 7th layer 711 of the block, 7th layer 711 to 8th layer 712, 8th layer 712 to The ninth layer 411 of the fourth cell 400, the tenth layer 412 of the ninth layer 411, the eleventh layer 511 of the fifth cell 500 of the tenth layer 412, The eleventh layer 511 is connected to the twelfth layer 512, the twelfth layer 512 to the thirteenth layer 611 of the sixth cell 600, and the thirteenth layer 611 to the fourteenth layer 612. An output value is output from the fourteenth layer 612.

As shown in Figure 1, once the connection structure is determined, such a connection structure is fixed so that there is no change in the structure. That is, the connection relationship between each cell is fixed, and the structure in the block may be changed.

As the connection relationship between the cells is fixed and the structure of the entire model is not deformed, the model structure can be maintained, thereby maintaining the performance of the model structure. Therefore, the model structure including the performance to be maintained can be maintained as the seed structure. For example, the model structure is a U-net, V-net structure, but is not limited to the above example.

In FIG. 1, for example, the first layer 111 is a 3X3X3 convolutional layer, the second layer 112 is a 5X5X5 maximum integrated layer, and the third layer 211 is a 5X5X5 convolutional layer. The fourth layer 212 is the 7X7X7 maximum integration layer, the fifth layer 311 is the 3X3 convolution layer, the sixth layer 312 is the 3X3X3 maximum integration layer, the seventh layer 711 is the 3X3 convolution layer, and the eighth layer Layer is the 3X3X3 maximum integration layer, the ninth layer 411 is a 5X5X5 convolution layer, the tenth layer 412 is a 3X3X3 maximum integration layer, the eleventh layer 511 is a 3X3X3 convolution layer, the twelfth layer 512 The 3X3X3 maximum integration layer, the thirteenth layer 611, the 3X3X3 convolutional layer, and the fourteenth layer 612 may be configured as the 3X3X3 maximum integration layer, but are not limited to the above example and may be configured in various layers.

2 to 4 are diagrams for explaining an example of an input and output relationship in which each cell is evolved in a learning system for reducing the amount of calculation according to an embodiment of the present invention.

First, a description will be given of a process in which blocks in each cell are evolved to form a model that provides optimal performance.

Each cell may contain N blocks, and each block may include up to M convolutional layers. The learning model consists of several cells, and all cells have the same internal structure.

An example of a process of constructing a block in which a block in a cell evolves to give an optimal performance will be described.

In the case where the nth output is generated by the first cell of the plurality of cells, the input that can enter the convolutional layer inside the block is the output of the n-1st cell, the output of the n-2nd cell, or the first. This is an output from inside the block belonging to the cell.

In one embodiment, the type of each convolution layer is a 1X1X1 convolution layer, a 3X3X3 convolution layer, a 5X5X5 convolution layer, a 7X7X7 convolution layer, a 3X3X3 maximum integration layer, a 3X3X3 average integration layer, a 3X3X3 atrus convolution layer ( atrous convolution layer) and 1X1X7 1X7X1 7X1X1 convolutional layer. The activation function is fixed with a rectified linear unit (ReLU).

According to the type of convolution layer, the number of newly added blocks is 8 * 8 * (2+ (M-1)) * (2+ (M-1)), and the number of starting models is 256 Dog. Each of the 256 models, the starting model, contains a dataset and the resulting value is measured. The result value includes a dice score or an accuracy. At this time, each model used for training is transformed into a unique expression that can be learned by the Recurrent Neural Network (RNN) by a conversion algorithm. Each of these expressions is input to RRN to perform regression analysis with the above result.

Then, M = 1, that is, all possible block combinations are used for models in which each block contains at most one convolutional layer, so that model candidates with M = 2 are generated. The number of model combinations created at this time is 256 * {8 ^ 2 * (2+ (2-1) ^ 2)} = 147,456. Then, the highest value of the predicted performances learned in the cyclic neural network is extracted and actually learned, and the actual performance and the model structure are again learned in the cyclic neural network.

The above process is repeated until M = 4, i.e., up to four convolutional layers of each block are included, thereby forming a block.

According to the block constructed by the above process, a model having an optimum performance is constructed.

2 to 4 are diagrams exemplarily illustrating a process of evolving each cell, which will be described sequentially from FIG. 2.

FIG. 2 is a diagram for describing a process of evolving a cell in the case of one block.

Referring to FIG. 2, the first cell 100 includes a first block 110, and the first block 110 includes a first layer 111, a second layer 112, and a first layer 111. And the sum 113 of the second layer 112.

When there is one block as shown in FIG. 2, the input values input to the first layer 111 and the second layer 112 included in the first block 110 become the N−1th output value 800. . The value output from the sum 113 of the first cell 100 is input to the N-th output 1000 via the concatenator 900. The Concat function is a function that concatenates strings.

FIG. 3 is a diagram for describing a process of evolving a cell in the case of two blocks.

Referring to FIG. 3, the first cell 100 includes a first block 110 and a second block 120, and the first block 110 includes the first layer 111 and the second layer 112. , The sum 113 of the first layer 111 and the second layer 112, and the second block 120 includes the third layer 121, the fourth layer 122, and the third layer 121. And the sum 123 of the fourth layer 122.

In the case of two blocks, as described above, the input that can enter the convolutional layer inside the block is the output of the n-1 th cell, the output of the n-2 th cell, or inside the block belonging to the first cell. Here is the output.

3, in the case of FIG. 3, the first layer 111, the second layer 112, and the third layer 121 are input with the N−1 th output value 800 as an input value. The input value of the layer 122 becomes the N-2nd output value 700. The sum 113 of the first layer 111 and the second layer 112 output from the first cell 100 and the third layer 121 and the fourth layer 122 output from the second cell 200. The output of the sum 123 is input to the N th output 1000 via the Concat 900.

FIG. 4 is a diagram for describing a process of evolving a cell in the case of three blocks.

Referring to FIG. 4, the first cell 100 includes a first block 110, a second block 120, and a third block 130, and the first block 110 includes the first layer 111. And a sum 113 of the second layer 112, the first layer 111, and the second layer 112, and the second block 120 includes the third layer 121 and the fourth layer 122. And a sum 123 of the third layer 121 and the fourth layer 122, and the third block 130 includes the fifth layer 131, the sixth layer 132, and the fifth layer 131. And the sum 133 of the sixth layer 132.

Even when there are three blocks, as described above, the input that can enter the convolutional layer inside the block is the output of the n-th cell, the output of the n-second cell, or the block belonging to the first cell. Is the output from

4, the N-2 output 700 is input to the first layer 111 of the first block 110, and the second block 120 is input to the second layer 112. The sum 123 of the sum of the values received by the third layer 121 and the fourth layer 122 from the N-th output 800 is input. In addition, the N−1 th output 800 is input to the fifth layer 131 of the third block 13 and the N−2 th output 700 is input to the sixth layer 132. The sum 113 of the first layer 111 and the second layer 112 output from the first cell 100, the third layer 121 and the fourth layer 122 output from the second cell 200. The sum 123 of the sum and the output of the sum 133 of the fifth layer 131 and the sixth layer 132 output from the third cell 300 are passed through a concating unit 900 and the Nth output. Inputted at 1000.

The above-described embodiments of FIGS. 2 to 4 exemplarily describe a process of evolving cells, and are not limited to the embodiments of FIGS. 2 to 4, and output the n-th output by the first cell of the plurality of cells. In this case, the input that can enter the convolutional layer inside the block is the output of the n-1th cell, the output of the n-2th cell, or the output from inside the block belonging to the first cell. All are included.

After only one cell of the plurality of cells has evolved, the number of blocks of all other cells is equally applied to the number of blocks of the one cell.

The learning system for reducing the amount of computation of the present invention is a network branch for a plurality of convolutional layers among a plurality of layers in order to solve the problem that a large amount of computation takes a long time when calculating a learning model based on a three-dimensional basis. The stroke can be performed.

In a network pruning process for a plurality of convolutional layers, all connections having a weight less than or equal to a predetermined value are deleted for each of the plurality of convolutional layers, and zero input weights are assigned to each of the plurality of convolutional layers. All one or more nodes with zero output weights are deleted. Subsequently, n + outputs the n-1th layer such that a node of the n-1th layer and the n + 1th layer of the same position where the node deleted from the nth layer among the plurality of convolutional layers is located are connected. This is done by adding a connection to the first layer.

The process of deleting all the connections having a weight less than or equal to a predetermined value is performed for each of the plurality of layers, but the deletion of each layer connection is sequentially performed. After all connections having a weight less than or equal to a predetermined value in the first layer are deleted, the first layer in which the connection is deleted is re-learned, and when the performance of the network is changed by re-learning, a predetermined value in the second layer is determined. All connections with the following weights are deleted.

That is, the process of deleting all the connections having a weight less than or equal to the predetermined value is repeated until all of the convolutional layers are fully implemented, and repeatedly having a weight less than or equal to a predetermined value for the connection of one layer. After the connection is completely deleted, the layer from which the connection is deleted is re-learned, and then all connections having a weight less than or equal to a predetermined value in the next layer are deleted.

In the above-described process of performing network pruning on the plurality of convolutional layers, 1 × 1 convolution is applied to each start convolutional node, thereby reducing the amount of computation of the convolution and improving performance by about 10%. In addition, the fully convolution network (FCN) is replaced by global average pooling, which saves about 7% of performance by saving computation and memory compared to the entire convolutional network before replacement. Is improved.

FIG. 5 illustrates a node of an n-1 th layer and n + 1 in the same position where a node deleted in an n th layer among a plurality of convolutional layers is located in a network pruning of a learning system that reduces an amount of computation according to an exemplary embodiment of the present invention. A diagram for describing a connection for sending an output of an n-1 th layer to an n + 1 th layer so that nodes of a first layer are connected.

Referring to FIG. 5, a node 811 of an n−1th layer, which is disposed at the same position, for a node represented by a dotted line in an nth layer 1011 deleted by network pruning among a plurality of convolutional layers. ) Is connected as the node 1211 of the n + 1th layer.

In the learning system for reducing the amount of computation of the present invention, a loss function combined with a loss function for a plurality of layers in order to increase the accuracy of image segmentation when the image is segmented using the learning model is combined. Calculated as a function).

The loss function is modified and calculated as a combined loss function, in which the similarity between the first region and the second region is calculated as a mixed matrix, and a dice value is measured, and the measured loss values are mixed to combine the loss function. It is modified as a function and by class balancing through weighted cross entropy.

The first area is an area extracted from the original image, and the second area is an area extracted from the label image, and the first area and the second area are areas of the same size in each of the original image and the label image.

In general, cross entropy is calculated as follows.

When calculating the cross entropy, if the ratio of the region of interest is unbalanced, the probability value of a particular class is lowered, so that the loss cannot be properly reflected in the learning model. Therefore, the accuracy of image segmentation can be increased only by class balancing through weighted cross entropy.

For the cross-entropy loss function and the die loss function, the prediction probability for each class i at each pixel x is called p (x, i) by applying softmax activation pixel by pixel in the final layer when training the model. When the target at position x is t (x), the cross-entropy loss function and the die loss function are defined as follows.

In the case of the cross-entropy loss function, the segmentation operation applies a standard entropy loss function to measure the pixel-by-pixel probability error between the predicted output and the target and sum the errors at all pixels. In addition, by applying a weight to each class in order to set an imbalance of pixel frequencies for each class, the weight for class i is w _i .

The weighted cross-entropy loss L _ce is expressed by Equation 1 below.

<Equation 1>

In the case of a die loss function, the dice's coefficient is a measure of the similarity between two samples. Extending the dice coefficient to the loss function improves performance when dealing with situations where the background pixel is higher than the label. Like weighted cross entropy, we use weights to set the class imbalance.

The dice loss L _dice is a sum of weights of the dice losses L _i for each class i, as shown in Equation 2 below.

<Equation 2>

In Equation 2, a binary map with respect to class i is represented by t _i, as shown in Equation 3 below.

<Equation 3>

Therefore, the die loss for class i is expressed by Equation 4 below.

<Equation 4>

As a result, the combined loss function is a function that satisfies Equation 5 below.

<Equation 5>

In Equation 5, L _{CE + dice} is a loss value reflecting cross entropy and dice value, dice is a value obtained by calculating the similarity between the first area and the second area as a mixed matrix, and λ _CE is a weight for the cross entropy loss function. The parameter, λ _dice, is a weighting parameter for the die loss function, and L _dice is a loss value that reflects the dice value.

FIG. 6 is a diagram illustrating a result of correcting a loss function of a learning system that reduces a computation amount according to an embodiment of the present invention and reflecting it on a medical image.

FIG. 6A is a diagram illustrating a segmentation of the liver in the medical image before the correction of the loss function, and FIG. 7B is a diagram illustrating the segmentation of the liver in the medical image after the correction of the loss function.

6 (a) is not divided correctly with respect to the pointed part of the liver or the parts that are not easily distinguished from the surrounding organs, but (b) of FIG. 7 is a part that is difficult to distinguish from the pointed part of the liver or the surrounding organs The results of the correct division can also be confirmed.

7 is a graph showing the result of the learning system to reduce the amount of calculation according to an embodiment of the present invention.

The graph of FIG. 7 is a result of measuring performance while evolving generation using a combined loss function while fixing the connection relationship between cells so that the aforementioned structural change does not occur.

Heuristic is the result of measuring the performance using the existing model. In the case of the existing model, each model is trained once and then the optimum value is produced.

Genetic is a result of measuring the performance by evolving generation using a combined loss function while fixing the connection between cells so that the above-mentioned structural change does not occur. Rather, only some models are trained after training to produce optimal values.

Referring to the graph of FIG. 7, it can be seen that in the case of genetics, the optimal value comes out when using less generations than the heuristic, although a much smaller model shows a learned result.

Therefore, in the case of using the combined loss function while fixing the connection relationship between the cells so that structural changes do not occur, there is an effect that the optimal performance and the optimal value can be quickly derived even with a small amount of computation.

The steps of a method or algorithm described in connection with an embodiment of the present invention may be implemented directly in hardware, in a software module executed by hardware, or by a combination thereof. Software modules may include random access memory (RAM), read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disk, removable disk, CD-ROM, or It may reside in any form of computer readable recording medium well known in the art.

In the above, embodiments of the present invention have been described with reference to the accompanying drawings, but those skilled in the art to which the present invention pertains may realize the present invention in other specific forms without changing the technical spirit or essential features thereof. I can understand that. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

100, 200, 300: cell
Block: 110, 120, 130, 210, 220, 230, 310, 320, 330
Layers: 111, 112, 121, 122, 131, 132, 211, 212, 221, 222, 231, 232, 311, 312, 321, 322, 331, 332

Claims (9)

In a learning system that reduces association,
Includes a plurality of cells including a plurality of blocks,
It is characterized in that the connection between the input and output between each cell is fixed to maintain a specific overall system structure,
The block includes a plurality of layers for performing input, processing or output of data,
The plurality of blocks or layers,
Characterized in that the structure can be changed in the data learning process,
Learning system that reduces the amount of computation. The method of claim 1,
In the case of outputting the nth output by the first cell of the plurality of cells,
An input that can enter the convolutional layer inside the block is
Characterized in that the output of the n-1 th cell, the output of the n-2 th cell or the output from the block belonging to the first cell,
Learning system that reduces the amount of computation. The method of claim 1,
After only one cell of the plurality of cells has evolved, the number of blocks of all other cells is applied equally to the number of blocks of the one cell.
Learning system that reduces the amount of computation. The method of claim 1,
Characterized in that the amount of calculation is reduced through network pruning for a plurality of convolutional layers of the plurality of layers,
Learning system that reduces the amount of computation. The method of claim 4, wherein
Network pruning for the plurality of convolutional layers,
For each of the plurality of convolutional layers, a connection having a weight less than or equal to a predetermined value is all deleted.
One or more nodes having zero input weights or zero output weights are all deleted for each of the plurality of convolutional layers.
N + 1 output of the n-1th layer such that the node of the n-1th layer and the n + 1th layer of the same position where the node deleted in the nth layer among the plurality of convolutional layers is located are connected; Characterized in that the connection to the second layer is added,
Learning system that reduces the amount of computation. The method of claim 5,
Deleting all connections having a weight less than or equal to the predetermined value
All are performed on each of the plurality of layers, and the deletion of each layer connection is sequentially performed.
In the first layer, all connections with weights less than or equal to a predetermined value are deleted.
Re-learning using the first layer where the connection is deleted,
When the performance of the network is changed by the re-learning, the connection having a weight less than or equal to a predetermined value in the second layer is all deleted.
Learning system that reduces the amount of computation. The method of claim 1,
For the plurality of layers, the accuracy of image segmentation is increased by modifying and calculating the loss function as a combined loss function.
Learning system that reduces the amount of computation. The method of claim 7, wherein
The loss function is modified to be calculated as a combined loss function,
The similarity between the first region and the second region is calculated with the mixed matrix, and the dice value is measured,
The dice values measured are mixed to modify the loss function as a combined loss function,
Characterized by class balancing through weighted cross entropy,
The first area is an area extracted from the original image, the second area is an area extracted from the label image,
Wherein the first area and the second area are areas of the same size in the original image and the label image, respectively.
Learning system that reduces the amount of computation. The method of claim 8,
The combined loss function is characterized by satisfying the following Equation 1,
Learning system that reduces the amount of computation.
<Equation 1>

In Equation 1, L _{CE + dice} is a loss value reflecting cross entropy and a dice value, dice is a value measured by calculating a similarity between the first area and the second area with a mixed matrix, and λ _CE is a value for the cross entropy loss function. The weighting parameter, λ _dice, is a weighting parameter for the die loss function, and L _dice is a loss value that reflects the _dice value.

Images