KR20220049855A - Apparatus and method for controlling distributed neural network - Google Patents

Abstract

Provided is a method for controlling a distributed neural network comprising: a step of classifying a plurality of filters into fewer parts than that of the number of student networks, when a plurality of groups comprising the plurality of filters are received; a step of transmitting the information regarding the same part to two or more student networks, to train the two or more student networks using one part; a step of receiving the results of the two or more student networks trained through the same part; and a step of generating a final feature based on an average value of the received results of the two or more student networks. Therefore, the present invention is capable of maintaining a constant inference accuracy.

Description

Control method and apparatus of a distributed neural network {APPARATUS AND METHOD FOR CONTROLLING DISTRIBUTED NEURAL NETWORK}

The present disclosure relates to a distributed neural network, and more particularly, to a method and apparatus for controlling a distributed neural network using a device having a low data rate.

Recently, with the spread of various Internet of Things (IoT) devices, various wearable devices such as smart watches and smart goggles are being developed. These wearable devices have low data transmission rates such as human body communication or Bluetooth communication using a human body as a medium due to low power requirements.

Meanwhile, with the development of artificial intelligence technology, the necessity of running neural network-based applications such as image classification and voice recognition even in devices with low data rates such as the aforementioned wearable devices is emerging.

In order to run a neural network-based application in a device with a low data rate, a technique for driving a neural network in an offloading manner has been developed. This is a method in which data is transmitted from the device to a cloud system using hardware such as a high-performance CPU, GPU, and accelerator, and a neural network is performed remotely, and then the data processing result is transmitted from the cloud system to the device again.

However, this offloading driving method may degrade the real-time performance because the neural network reasoning speed may be lowered depending on the remote server state and the network communication state. In addition, as the number of devices increases, frequent communication may occur, which may cause additional delay and increase power consumption of the devices. Accordingly, there is a need for a method of mounting and driving a neural network in a device having a low data rate.

An object of the present embodiment is to provide a method and apparatus for controlling a distributed neural network, in which a student network is mounted on a device having a low data rate.

The technical problems to be achieved by the present embodiment are not limited to the technical problems described above, and other technical problems may be inferred from the following embodiments.

According to a first embodiment, a method for controlling a distributed neural network includes, when a plurality of groups including a plurality of filters are received, classifying the plurality of filters into a number of parts smaller than the number of student networks, one part transmitting information about the same part to the two or more student networks, receiving results of the two or more student networks learned through the same part, and the received two or more student networks to train two or more student networks using generating the final feature based on the average value of the results of the student network.

According to the second embodiment, when a plurality of groups including a plurality of filters are received by executing a memory for storing at least one instruction and the at least one instruction, In order to classify the plurality of filters into a number of parts less than the number of student networks, and to train two or more student networks using one part, information about the same part is transmitted to the two or more student networks, and the same and a processor configured to receive a result of the two or more student networks learned through the part, and generate a final feature based on an average value of the received results of the two or more student networks.

According to a third embodiment, a computer-readable recording medium is a computer-readable non-transitory recording medium recording a program for executing a method for controlling a distributed neural network in a computer, the method for controlling a distributed neural network comprising: When a plurality of groups including a plurality of filters are received, classifying the plurality of filters into a number of parts less than the number of student networks, to train two or more student networks using one part, the two or more students transmitting information about the same part to a network; receiving results of two or more student networks learned through the same part; and generating a final feature based on an average value of the received results of two or more student networks. may include

Specific details of other embodiments are included in the detailed description and drawings.

The method and apparatus for controlling a distributed neural network according to the present disclosure have the effect of maintaining constant reasoning accuracy even if communication of some of the devices on which the student network is mounted is unstable.

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

1 is a diagram for explaining an example of a distributed neural network structure.
2A, 2B, and 3 are diagrams for explaining an example of a network of neural network (NONN)-based distributed neural network structure.
4 is a diagram for describing a structure of a distributed neural network according to an embodiment.
5 is a flowchart illustrating a distributed neural network control method according to an embodiment.
6A and 6B are diagrams for explaining a distributed neural network control method according to another embodiment.
7A, 7B, 8A, and 8B are diagrams for explaining the performance of a distributed neural network according to an embodiment.
9 is a block diagram illustrating an apparatus for controlling a distributed neural network according to an embodiment.

Terms used in the embodiments are selected as currently widely used general terms as possible while considering functions in the present disclosure, but may vary according to intentions or precedents of those of ordinary skill in the art, emergence of new technologies, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the corresponding description. Therefore, the terms used in the present disclosure should be defined based on the meaning of the term and the contents of the present disclosure, rather than the simple name of the term.

In the entire specification, when a part "includes" a certain element, this means that other elements may be further included, rather than excluding other elements, unless otherwise stated.

The expression "at least one of a, b, and c" described throughout the specification means 'a alone', 'b alone', 'c alone', 'a and b', 'a and c', 'b and c ', or 'all of a, b, and c'.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art to which the present disclosure pertains can easily implement them. However, the present disclosure may be implemented in several different forms and is not limited to the embodiments described herein.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings.

1 is a diagram for explaining an example of a distributed neural network structure.

Referring to FIG. 1 , the neural network 100 extracts knowledge through a filter to generate a single student network 120 that learns the final result of the teacher network 110 with a large memory requirement and large computational amount. Here, the teacher network 110 may be a relatively high-performance network, and the student network 120 may be referred to as a sub-network that learns using some knowledge of the teacher network 110 .

However, in the case of the student network 120 of FIG. 1 , it may be difficult to operate in a single device having a low data transfer rate because the required amount of computation and memory capacity are still large. Accordingly, the student network 120 may be distributed and mounted on several devices 131 to 133 on a layer-by-layer basis. In this case, the second device 132 may learn by using some knowledge of the student network 120 and an inference result of the first device 131 . Also, the third device 133 may learn based on some knowledge of the student network 120 and an inference result of the second device 132 and generate a final feature map.

That is, in the distributed neural network 100 of FIG. 1 , data transmission and reception between all devices 131 to 133 is inevitable in order to obtain a single reasoning result, and thus a required communication amount and communication frequency between devices may be high. Also, when communication of any one device among the first device 131 to the third device 133 is interrupted, it may be impossible to derive an inference result. Accordingly, it may be difficult to mount the distributed neural network 100 of FIG. 1 in a device having a low data rate, such as a wearable device.

2A, 2B, and 3 are diagrams for explaining an example of a structure of a NONN-based distributed neural network.

As another example of the distributed neural network, there is a NONN (Network of Neural Network) based distributed neural network. Referring to FIG. 2A , the NONN-based distributed neural network 200 may include a teacher network 210 and a plurality of student networks 221 to 223 . In addition, the plurality of student networks 221 to 223 may be trained using each independent filter included in the teacher network 210 , and the plurality of student networks 221 to 223 may each have separate devices 231 to 233 . ) can be mounted on The final reasoning result of the distributed neural network 200 may be generated based on the respective outputs of the separate devices 231 to 233 .

Meanwhile, a method of transferring knowledge from the teacher network 210 to the plurality of student networks 221 to 223 may be described with reference to FIG. 2B . First, the filter 250 included in the final convolutional layer 240 of the teacher network 210 having a large amount of memory and computation may be classified into a plurality of groups. Thereafter, the filters 250 included in the plurality of groups may be classified into a plurality of parts again. The NONN-based distributed neural network 200 may transmit each filter 250 classified into a plurality of parts to a plurality of student networks (a first student network and a second student network). Here, the plurality of student networks (the first student network and the second student network) may all be independent. Thereafter, the NONN-based distributed neural network 200 may derive a result by concatenating the final results of each student network (the first student network and the second student network). Meanwhile, although it is shown that a plurality of groups and a plurality of parts are generated in the teacher network 210 in FIG. 2B , it is to those skilled in the art that the places where the plurality of groups and the plurality of parts are generated may vary. self-evident

The NONN-based distributed neural network 200 can provide high inference accuracy because it utilizes all the knowledge of the teacher network 210 when the communication of the device is stable. In addition, when the communication of the device on which the student network is mounted is unstable, it is possible to derive a result unlike the distributed neural network 100 of FIG. 1 . However, when it is difficult to receive the result learned by the communication-disrupted device, the NONN-based distributed neural network 200 uses only some knowledge of the teacher network 210 , so the inference accuracy may be greatly reduced.

On the other hand, the NONN-based distributed neural network 200 may classify the filters 250 included in the final convolutional layer 240 into a plurality of groups based on the AH (Activation Hub) rule in order to improve inference accuracy. .

According to an embodiment, the NoNN-based distributed neural network 200 uses an AH rule so that knowledge extracted through the filter 250 included in the final convolutional layer 240 of the teacher network 210 is not biased to a specific group. Filters can be classified into a plurality of groups based on The AH rule is a rule for classifying the filter 250 having great importance with respect to the same type of input. The average of the output of the i-th filter in the final convolutional layer 240 of the teacher network 210 is a _i , the j-th filter. When the output average of is a _j , the AH rule can be expressed as in Equation 1.

[Equation 1]

In order to classify the filters 250 of the final convolutional layer 240 into groups, each filter 250 is a vertex, and the weight of the edges between the filters 250 is a graph type having an AH value. data structures can be created. Later, AH values and Newman, Mark EJ. "Modularity and community structure in networks." Based on the segmentation algorithm disclosed in Proceedings of the national academy of sciences 103.23 (2006): 8577-8582., the filter 250 may be classified into a plurality of groups, and this may be referred to as a community structure. .

Based on the AH rule, when each filter 250 is classified into a plurality of groups, filters having a large output value for the same input may be classified into different groups because the AH value is small. For example, when there are several filters capable of discriminating dogs from objects in an image, each filter may be classified into a different group. This process can also be seen as a process of dispersing filters with high importance.

Meanwhile, since the number of filters included in each group classified by the AH rule may be different, the NoNN-based distributed neural network 200 provides a plurality of groups (first to fourth groups) as shown in FIG. 2B. The included filters may be classified into parts (a first part and a second part) again. In this case, the filters may be classified so that the number of filters included in each part may be uniform. Thereafter, the NoNN-based distributed neural network 200 may train the student network using each part.

On the other hand, in the NoNN-based distributed neural network, when the device on which the student network is mounted fails to communicate, the results of the student network learned through some filters of the teacher network cannot be used, so inference accuracy may be reduced.

Meanwhile, FIG. 3 is a diagram for explaining a NONN-based distributed neural network that trains four student networks.

If the filters included in the final convolutional layer of the teacher network are classified into groups 1 to 5 as shown in FIG. 3 and the number of devices on which the student network is to be mounted is 4, the NONN-based distributed neural network is Filters can be classified into the same number of parts as the number of networks. The number of filters included in each group, part and student network of FIG. 3 is shown in parentheses.

Meanwhile, in order to make the number of filters included in each part as uniform as possible, the filters (7) included in the fourth group and the filters (5) included in the fifth group may be classified into the fourth part.

If an error occurs in the communication of the device equipped with the first student network in the NoNN-based distributed neural network configured as described in FIG. 3, since the result learned based on the first part cannot be used, the final feature map (Final Elements 0 to 9 of the feature map 310 have a value of 0. Accordingly, inference accuracy may be deteriorated.

On the other hand, since the wearable device has a lot of movement of the human body and uses wireless communication or human body communication, the communication state may be unstable. Therefore, in order for the student network to be mounted on devices with low data rates, including wearable devices, it is necessary to control the distributed neural network so that the distributed neural network can be robust to communication errors.

4 is a diagram for describing a structure of a distributed neural network according to an embodiment.

The distributed neural network according to an embodiment may be based on NONN. Accordingly, the process of classifying the filters included in the final convolutional layer of the teacher network into the first to fifth groups may correspond to the classification process of FIGS. 2B and 3 .

Meanwhile, in the case of a distributed neural network according to an embodiment, the filter may be classified into a number of parts smaller than the number of student networks. For example, when the number of devices and the number of student networks to be mounted on the device are four, the number of parts may be two less than four. Filters included in the first to fifth groups may be numbered as uniformly as possible in the first part and the second part. Accordingly, the filters included in the first group and the third group may be classified as the first part, and the filters included in the second group, the fourth group, and the fifth group may be classified as the second part. If the number of parts is less than the number of devices, the number of student networks learning using one part becomes a large number, so even if any one device loses communication, the final feature map can be created as a result of learning using the filters of all parts. can

Referring to FIG. 4 , the student network learning using the first part may be a first student network and a third student network, and the student network learning using the second part is a second student network and a fourth student network can be

Thereafter, the distributed neural network control method may calculate the average of the results of the student network learned using the same part. Referring to FIG. 4 , a first part average that is an average value of the results of the first student network and the third student network may be calculated, and a second part average that is an average value of the results of the second student network and the fourth student network is calculated can be

Finally, the distributed neural network control method may generate a final feature map by concatenating the first part average and the second part average. In this case, even if a communication error occurs in the device equipped with the first student network, the final feature map can be generated by reflecting the learning result using the filter included in the first part through the third student network, so the communication error The change in the final feature map due to .

5 is a flowchart illustrating a method for controlling a distributed neural network according to an embodiment.

In step S510 , the distributed neural network control method may classify the plurality of filters into a number of parts smaller than the number of student networks when a plurality of groups including the plurality of filters are received. Here, the number of parts may be determined based on a value obtained by dividing the number of student networks by 2. The distributed neural network control method according to an embodiment may determine the number of parts based on Equation (2).

[Equation 2]

은 내림(버림) 값을 의미한다. 예를 들어, 학생 네트워크가 탑재될 디바이스의 수가 5개인 경우, 학생 네트워크의 수(M)는 5이고, 파트의 수(N)는 2.5의 내림 값인 2 이하일 수 있다. 다시 말해, 학생 네트워크가 5개이면 파트의 수는 1개 또는 2개일 수 있다. Here, N means the number of parts, and M means the number of student networks.

stands for a rounded down (truncated) value. For example, when the number of devices on which the student network is to be mounted is 5, the number of student networks (M) may be 5, and the number of parts (N) may be 2 or less, which is a rounded-down value of 2.5. In other words, if there are 5 student networks, the number of parts may be 1 or 2.

Meanwhile, when the number of student networks learned through each part is different, the number of student networks learned through each part may be determined based on the number of filters included in each part. According to an embodiment, in the method of controlling a distributed neural network including N parts and M devices, the number (S _i ) of the student network to be learned using the part including the i-th most filters among the parts It can be obtained through Equation 3.

[Equation 3]

For example, the number of student networks (M) and the number of parts (N) are 5 and 2, respectively, the number of filters included in the first part is 30, and the number of filters included in the second part is 29. there is. At this time, if i=1, the part containing more filters among the first part and the second part is the first part, so the number of student networks to be learned using the first part (S ₁ ) is 5/2 It may be a rounded value of , that is, three. And if i=2, the number of student networks learning using the second part (S ₂ ) is 5/ It can be a rounding-down value of 2, that is, two. Accordingly, the number of student networks learning through the first part may be three, and the number of student networks learning through the second part may be two.

In addition, step S510 may be to classify the number of filters classified into each part to be uniform.

In step S520, the distributed neural network control method may transmit information about the same part to two or more student networks in order to train two or more student networks using one part.

In step S530, the distributed neural network control method may receive the results of two or more student networks learned through the same part.

In step S540, the distributed neural network control method may generate a final feature based on an average value of the received results of two or more student networks. The method for controlling a distributed network according to an embodiment may generate a final feature map by connecting average values of results of the student network.

Therefore, the control method of the distributed network of the present disclosure may maintain inference accuracy even if the results of some student networks are not transmitted due to communication failure of a specific device by reinforcing redundancy between parts.

6A and 6B are diagrams for explaining a distributed neural network control method according to another embodiment.

A method of controlling a distributed neural network according to another embodiment may determine a student network to be loaded in each device based on information about communication stability of each device. In this case, the control method of the distributed neural network may determine a device on which the student network is to be mounted based on information about the overall reasoning accuracy of the neural network when the communication of each student network fails.

Referring to FIG. 6A , the first to fifth student networks may be mounted on one of the first to fifth devices, respectively. The first list 610 of FIG. 6A is a list in which inference accuracy is high when each student network is used independently, and the second list 620 is a list in the order in which communication stability of each device is high. In other words, the third student network is a student network that can provide higher reasoning accuracy compared to the case of using other student networks independently, and the first device may be a device that provides the most stable data transmission and reception among devices. On the other hand, the inference accuracy when using the student network independently can be obtained through the experimental results shown in FIGS. 7B and 8B, and information on the communication stability of each device can be obtained from the device manufacturer or testing institution, but limited to this doesn't happen

In the method of controlling a distributed neural network according to an embodiment, each student network may be mounted on each device according to the order of the first list 610 and the order of the second list 620 . In this case, since the student network that provides the highest inference accuracy is mounted on the device that provides the most stable data transmission and reception, deterioration of inference accuracy according to the communication state can be reduced.

In addition, the method for controlling a distributed neural network according to another embodiment may determine a device on which the student network is to be mounted based on a part used to learn the student network.

Referring to FIG. 6B , the third list 630 is a list in the order of highest inference accuracy when the student network learned based on the first part is used independently, and the fourth list 640 is the second part This is a list in order of highest inference accuracy when independently using the student network trained on the basis of In this case, it may be determined to load each device included in the second list 620 with the student network learned through the first part and the student network learned through the second part cross each other. For example, it may be determined to mount the first student network learned through the first part on the first device, and it may be determined to mount the second student network learned through the second part on the second device.

That is, in the method of controlling a distributed neural network according to an embodiment, when the student networks learned based on the n-th part are independently used, the x-th element of the list sorted in the order of high inference accuracy is P _n,x When doing so, the set of student networks to be mounted on the i-th device having the highest communication stability (D _i ) may be determined based on Equation (4).

[Equation 4]

According to Equation 4, the set (D ₁ ) of the student network to be mounted on the device with the highest communication stability includes the student network (P _1,1 ) having the highest inference accuracy among the networks learned based on the first part. can If the device on which the student network is to be mounted is determined based on these criteria, the probability of simultaneous communication errors occurring in the student network learned based on the same part is minimized, so even in an environment with poor communication conditions, the inference accuracy of the distributed neural network is improved. The degree of degradation may be reduced.

That is, even if a communication error occurs in some devices and only results from some student networks are used, inference with similar accuracy to that of the existing teacher network is possible. In addition, in the case of a distributed neural network according to an embodiment, since only the final result of each student network is transmitted, it can be mounted on a device having a low data transmission rate, including a wearable device.

7A, 7B, 8A, and 8B are diagrams for explaining the performance of a distributed neural network according to an embodiment.

In order to check the performance of the distributed neural network control method according to an embodiment, the conventional distributed neural network based on NONN and the distributed neural network according to the embodiment were trained using eight student networks, respectively. Then, for each of the 8 devices equipped with the 8 student networks, when a communication error occurred, the output value of the corresponding device was set to 0, and inference accuracy was measured. On the other hand, an experiment assuming a communication error was conducted for all possible cases in 8 student networks.

7A is a graph showing the inference accuracy of a conventional distributed neural network based on NONN using the Cifar10 dataset, and FIG. 7B is a graph showing the inference accuracy of the distributed neural network according to an embodiment using the Cifar10 dataset. A teacher network included in the distributed neural network of FIGS. 7A and 7B is WRN40-4.

7A and 7B, when the number of devices that have failed to communicate when using the Cifar10 dataset is three, the distributed neural network according to an embodiment has an average inference accuracy of 1.164%, the lowest compared to the conventional distributed neural network based on NONN. It can be seen that the inference accuracy is improved by 11.22%. This improvement becomes more evident as the number of devices that have failed to communicate increases. 7A and 7B, when the number of devices that have failed communication is 7, the distributed neural network according to an embodiment has improved average inference accuracy by 18.741% and lowest inference accuracy by 54.22% compared to the conventional distributed neural network based on NONN. can be checked

Meanwhile, FIG. 8A is a graph showing the inference accuracy of a NONN-based distributed neural network using the Cifar100 dataset, and FIG. 8B is a graph showing the inference accuracy of a distributed neural network using the Cifar100 dataset according to an embodiment. A teacher network included in the distributed neural network of FIGS. 8A and 8B is WRN28-10.

Referring to FIGS. 8A and 8B , when the number of devices that have failed communication when using the Cifar100 dataset is three, the distributed neural network according to an embodiment has improved average inference accuracy by 15.072% compared to the conventional distributed neural network based on NONN. can check that In addition, when the number of devices that have failed communication is 7, it can be seen that the average inference accuracy of the distributed neural network according to an embodiment is improved by 18.742% compared to the conventional distributed neural network based on NONN.

7A, 7B, 8A and 8B , in all cases where the distributed neural network according to an embodiment uses the Cifar10 dataset and the Cifar100 dataset, the communication failure of the device compared to the conventional NoNN-based distributed neural network You can see it's strong.

9 is a block diagram illustrating an apparatus for controlling a distributed neural network according to an embodiment.

The distributed neural network control apparatus 900 may include a memory 910 and a processor 920 , according to an embodiment. In the distributed neural network control apparatus 900 shown in FIG. 9, only the components related to this embodiment are shown. Therefore, it can be understood by those of ordinary skill in the art related to the present embodiment that other general-purpose components may be further included in addition to the components shown in FIG. 9 .

The memory 910 is hardware for storing various types of data processed in the distributed neural network control apparatus 900 . For example, the memory 910 includes data processed in the distributed neural network control apparatus 900 and to be processed. data can be stored. The memory 910 may store at least one instruction for an operation of the processor 920 . Also, the memory 910 may store a program or an application to be driven by the distributed neural network control apparatus 900 . The memory 910 is a random access memory (RAM), such as dynamic random access memory (DRAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), CD- It may include ROM, Blu-ray or other optical disk storage, a hard disk drive (HDD), a solid state drive (SSD), or flash memory.

The processor 920 may control the overall operation of the distributed neural network control apparatus 900 and process data and signals. The processor 920 may generally control the distributed neural network control apparatus 900 by executing at least one instruction or at least one program stored in the memory 910 . The processor 920 may be implemented as a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), or the like, but is not limited thereto.

When a plurality of groups including a plurality of filters are received, the processor 920 classifies the plurality of filters into a number of parts less than the number of the student networks, and uses one part to train two or more student networks. It is possible to transmit information about the same part to more than one student network, receive results of two or more student networks learned through the same part, and generate a final feature based on an average value of the received results of two or more student networks.

In this case, the number of parts may be determined based on a value obtained by dividing the number of student networks by 2, and when the number of student networks learned through each part is different, each part is selected based on the number of filters included in each part. The number of student networks to be learned through may be determined.

Meanwhile, the processor 920 may determine a device on which the student network will be mounted, based on information about the communication stability of the device on which the student network will be mounted. In this case, the processor 920 may determine a device on which the student network is to be mounted, based on information about the overall reasoning accuracy of the neural network when the communication of each student network fails. Also, the processor 920 may determine a device on which the student network is to be mounted, based on the part used to learn the student network.

Meanwhile, when classifying the plurality of filters into parts, the processor 920 may classify the filters so that the number of filters classified into each part is uniform. Also, the distributed neural network may be based on NONN.

The processor according to the above-described embodiments includes a processor, a memory for storing and executing program data, a permanent storage such as a disk drive, a communication port for communicating with an external device, a touch panel, a key, a button, etc. user interface devices, and the like. Methods implemented as software modules or algorithms may be stored on a computer-readable recording medium as computer-readable codes or program instructions executable on the processor. Here, the computer-readable recording medium includes a magnetic storage medium (eg, read-only memory (ROM), random-access memory (RAM), floppy disk, hard disk, etc.) and an optically readable medium (eg, CD-ROM). ), and DVD (Digital Versatile Disc)). The computer-readable recording medium may be distributed among network-connected computer systems, so that the computer-readable code may be stored and executed in a distributed manner. The medium may be readable by a computer, stored in a memory, and executed on a processor.

This embodiment may be represented by functional block configurations and various processing steps. These functional blocks may be implemented in any number of hardware and/or software configurations that perform specific functions. For example, an embodiment may be an integrated circuit configuration, such as memory, processing, logic, look-up table, etc., capable of executing various functions by the control of one or more microprocessors or other control devices. can be hired Similar to how components may be implemented as software programming or software components, this embodiment includes various algorithms implemented in a combination of data structures, processes, routines or other programming constructs, including C, C++, Java ( Java), assembler, etc. may be implemented in a programming or scripting language. Functional aspects may be implemented in an algorithm running on one or more processors. In addition, the present embodiment may employ the prior art for electronic environment setting, signal processing, and/or data processing. Terms such as “mechanism”, “element”, “means” and “configuration” may be used broadly and are not limited to mechanical and physical configurations. The term may include the meaning of a series of routines of software in connection with a processor or the like.

The above-described embodiments are merely examples, and other embodiments may be implemented within the scope of the claims to be described later.

Claims (10)

when a plurality of groups including a plurality of filters are received, classifying the plurality of filters into a number of parts less than the number of student networks;
In order to train two or more student networks using one part, transmitting information about the same part to the two or more student networks;
receiving results of two or more student networks learned through the same part; and
and generating a final feature based on an average value of the received results of the two or more student networks. According to claim 1,
The method of controlling a distributed neural network, wherein the number of parts is determined based on a value obtained by dividing the number of the student network by 2. According to claim 1,
If the number of student networks taught through each part is different,
Based on the number of filters included in each part, the number of student networks learned through each part is determined, the control method of a distributed neural network. According to claim 1,
Based on the information about the communication stability of the device, the method further comprising the step of determining a device on which the student network will be mounted, the control method of a distributed neural network. 5. The method of claim 4,
The step of determining the device on which the student network is to be mounted,
When each student network fails to communicate, based on information about the overall neural network reasoning accuracy, a device to which the student network is mounted is determined, a control method of a distributed neural network. 5. The method of claim 4,
The step of determining the device on which the student network is to be mounted,
A method of controlling a distributed neural network, which is based on the part used to train the student network. According to claim 1,
The step of classifying the plurality of filters is
A method of controlling a distributed neural network that classifies the number of filters classified into each part to be uniform. According to claim 1,
The distributed neural network is a method of controlling a distributed neural network, which is based on a Network of Neural Network (NONN). a memory for storing at least one instruction; and
By executing the at least one command,
When a plurality of groups including a plurality of filters are received, classifying the plurality of filters into a number of parts less than the number of student networks,
In order to train two or more student networks using one part, sending information about the same part to the two or more student networks,
receiving the results of two or more student networks trained through the same part;
and a processor for generating a final feature based on an average value of the received results of the two or more student networks. As a computer-readable non-transitory recording medium in which a program for executing a control method of a distributed neural network is recorded in a computer,
The control method of the distributed neural network,
when a plurality of groups including a plurality of filters are received, classifying the plurality of filters into a number of parts less than the number of student networks;
In order to train two or more student networks using one part, transmitting information about the same part to the two or more student networks;
receiving results of two or more student networks learned through the same part; and
and generating a final feature based on an average value of the received results of the two or more student networks.

Images