KR102096737B1 - Distributed machine learning method with fault tolerance using LDPC codes and apparatus therefore - Google Patents

Abstract

Disclosed are a distributed machine learning method with a failure prevention ability using low density parity check (LDPC) codes and a device thereof. According to an embodiment of the present invention, a distributed machine learning error recovery method comprises the steps of: receiving calculation result values from workers; removing a variable node of an erasure and a check node connected to the variable node of the erasure among variable nodes; determining an error candidate among the variable nodes based on the calculation result value of the variable nodes connected to each of check nodes; and correcting the calculation result value of an error node connected to the corresponding check node based on the calculation result value of each of the variable nodes connected to the check node connected to the determined error candidate.

Description

Distributed machine learning method with fault tolerance using LDPC codes and apparatus therefore}

The present invention relates to a technique for recovering an error from a distributed machine learning in which a failure occurs, and more specifically, to a master in a distributed machine learning environment by using a decoding feature of a low density parity check (LDPC) code. A method and apparatus for detecting and correcting an incorrect calculation result in the case of sending an incorrect calculation result value.

In recent years, the paradigm for large-scale machine learning and data analysis has moved to large distributed systems. In particular, modern distributed systems and computational primitive systems have attracted much attention. Recently, a method for dividing and calculating a large-scale matrix calculation using a coding theory technique was introduced, and the introduced system is as follows.

에 포함되는 충분히 큰 행렬이고, x는

의 원소일 수 있다. 결과 값인 Ax를 얻기 위해서, 마스터는 행렬 A를 A₁, ?, A_k로 나눈다. 여기서,

와 (n, k) MDS(maximum distance separable) 부호를 이용하여 부분행렬(submatrix) A_i를 도 2에 도시된 바와 같이 B₁, ?, B_n로 부호화(encoding)할 수 있으며,

일 수 있다.First, the system presupposes n workers, and the desired value to get in a given situation is Ax. Where A is

Is a sufficiently large matrix, x is

It may be an element of. To obtain the result, Ax, the master divides matrix A by A ₁ ,?, A _k . here,

And (n, k) sub-matrix A _i using MDS (maximum distance separable) codes can be encoded into B ₁ ,?, B _n as shown in FIG. 2,

Can be

Now we send B _i and x to n workers, and n workers calculate B _i x and when the calculation is finished, we send a vector of length r / k which is the result of the calculation to the master. Even if a calculation result is received from k workers among n workers from the nature of the MDS code, Ax, a desired value, can be decoded. In other words, (nk) workers with a slow calculation rate are defined as a straggler, and when the calculation result is not received, the received information is treated as an erasure to perform decoding.

의 영공간(null space)으로 정의된다. 여기서, d_v와 d_c는 각각 패리티 검사 행렬에서 각 열과 행의 0이 아닌 성분, 즉 1의 개수를 의미할 수 있다.The results in the existing distributed machine learning study presuppose that workers can do the same amount of work in unit hours. In other words, the number of rows of the matrix A is divided into the same (homogeneous) size, encoded, and distributed to workers to perform the work. And the problem of distributing work by assuming that workers can do a heterogeneous amount of work for a unit time is solved in an asymptotic way. The error correction code used for decoding is a low density parity check (LDPC), and the (d _v , d _c ) -LDPC code is a parity check matrix H = [h _mn ] _{M × N}

It is defined as the null space of. Here, d _v and d _c may mean the number of non-zero components of each column and row in the parity check matrix, that is, 1, respectively.

However, in the existing distributed machine learning, it can be a problem to judge that all information received by the master is correct in many situations, such as when a worker sends malicious or wrong information to the master and the channel noise between the worker and the master is large. .

Embodiments of the present invention can detect and correct an incorrect calculation result when workers send an incorrect calculation result value to a master in a distributed machine learning environment by using a decoding feature of a low density parity check (LDPC) code. It provides a method and apparatus.

A distributed machine learning error recovery method according to an embodiment of the present invention comprises the steps of receiving calculation result values from workers; Removing a variable node of an erasure and a check node connected to the variable node of the erasure among variable nodes; Determining an error candidate among variable nodes based on a calculation result value of variable nodes connected to each of the inspection nodes; And correcting the calculation result value of the error node connected to the corresponding inspection node based on the calculation result value of each of the variable nodes connected to the inspection node connected to the determined error candidate.

The determining step calculates a first determination value for detecting the error candidate for each of the check nodes based on a calculation result value of variable nodes connected to each of the check nodes, and for each of the check nodes The error node may be determined based on the calculated first determination value.

In the determining step, for each of the variable nodes, when the first determination values of all the inspection nodes connected to the variable node are a preset value, the variable node may be determined as an error candidate.

In the correcting step, the error node connected to the first check node is calculated such that the sum of the calculation result values of each of the variable nodes connected to the first check node including only one of the determined error candidates is a preset value. The resulting value can be modified.

Furthermore, in the distributed machine learning error recovery method according to an embodiment of the present invention, after the calculation result value of the error node connected to the first inspection node is corrected, the determining step or the correcting step is repeated a predetermined number of times. It may further include the step of performing.

Furthermore, in the distributed machine learning error recovery method according to an embodiment of the present invention, after repeatedly performing the predetermined number of times, performing erasure decoding based on the calculation result values of variable nodes connected to each of the inspection nodes It may further include.

A distributed machine learning error recovery apparatus according to an embodiment of the present invention includes a receiver configured to receive calculation result values from workers; Among variable nodes, the variable node of erasure and the check node connected to the variable node of the eraser are removed, and the variable is based on the calculation result values of the variable nodes connected to each of the inspect nodes. A determination unit for determining error candidates among the nodes; And a correction unit for correcting the calculation result value of the error node connected to the corresponding inspection node based on the calculated result value of each of the variable nodes connected to the inspection node connected to the determined error candidate.

The determination unit calculates a first determination value for detecting the error candidate for each of the check nodes based on the calculation result values of the variable nodes connected to each of the check nodes, and is calculated for each of the check nodes The error node may be determined based on a first determination value.

The determining unit may determine the variable node as an error candidate when the first determination values of all the inspection nodes connected to the variable node are preset values for each of the variable nodes.

The correction unit calculates the result value of the error node connected to the first check node such that the sum of the calculation result values of each of the variable nodes connected to the first check node including only one of the determined error candidates is a preset value. Can be modified.

The determination unit and the correction unit may repeat the process of determining the error candidate and correcting the calculation result value for the error candidate a predetermined number of times after correcting the calculation result value of the error node connected to the first inspection node. have.

Furthermore, the distributed machine learning error recovery apparatus according to an embodiment of the present invention further repeats the predetermined number of times, and then further decodes the eraser decoding based on the calculation result values of the variable nodes connected to each of the inspection nodes. It can contain.

According to embodiments of the present invention, in the distributed machine learning environment, by using the decoding feature of the LDPC code, it is possible to detect and correct an incorrect calculation result when workers send an incorrect calculation result value to the master.

According to embodiments of the present invention, in a distributed machine learning environment, distributed machine learning may be operated in consideration of both a case where a worker is a straggler and a case where an incorrect calculation result value is transmitted.

According to embodiments of the present invention, errors that may occur in distributed machine learning can be efficiently detected and corrected using LDPC codes and bipartite graph representations of codes.

According to the embodiments of the present invention, errors that may occur in various forms in a machine learning field using a matrix multiplication of a tremendous size can be corrected, and since LDPC codes are used, they have lower complexity than other codes, and thus delay speed. It can also have good performance in parts.

Since the present invention corrects errors through the structure of LDPC codes in distributed machine learning, it can be useful even when channel noise is large in the process of transmitting distributed computing when malicious or accidental wrong information is transmitted. It can be used in a variety of forms in different fields of learning.

1 is a flowchart illustrating an operation of a distributed machine learning error recovery method according to an embodiment of the present invention.
2 to 5 show exemplary views for explaining the method of the present invention.
6 shows an exemplary diagram for evaluating the performance of the method of the present invention.
7 illustrates a configuration of a distributed machine learning error recovery apparatus according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them will be clarified with reference to embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various different forms, and only the present embodiments allow the disclosure of the present invention to be complete, and common knowledge in the art to which the present invention pertains. It is provided to completely inform the person having the scope of the invention, and the present invention is only defined by the scope of the claims.

The terminology used herein is for describing the embodiments, and is not intended to limit the present invention. In this specification, the singular form also includes the plural form unless otherwise specified in the phrase. As used herein, “comprises” and / or “comprising” refers to the components, steps, operations, and / or elements mentioned above of one or more other components, steps, operations, and / or elements. Presence or addition is not excluded.

Unless otherwise defined, all terms (including technical and scientific terms) used in the present specification may be used as meanings commonly understood by those skilled in the art to which the present invention pertains. In addition, terms defined in the commonly used dictionary are not ideally or excessively interpreted unless explicitly defined.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The same reference numerals are used for the same components in the drawings, and duplicate descriptions for the same components are omitted.

In embodiments of the present invention, it is intended to detect and correct an incorrect calculation result when workers send an incorrect calculation result value to a master in a distributed machine learning environment using a decoding feature of an LDPC code.

In this case, the present invention can operate distributed machine learning in consideration of both a case where a worker is a straggler and a result of incorrect calculation in a distributed machine learning environment. That is, the present invention can solve the problem in a situation where the straggler and the error coexist based on the structural characteristics of the LDPC code.

The present invention can efficiently detect and correct errors that may occur in distributed machine learning by using LDPC codes and bipartite graph representations of codes in order to efficiently detect and correct errors that may occur in distributed machine learning. have.

The method according to the present invention can be easily extended to an irregular LDPC code. The decoder of the LDPC code can be expressed by a Tanner graph composed of a variable node, a check node, and an edge.

Here, unlike the conventional LDPC code decoding method, a difference may exist in a method in which an unsatisfied test node is selected, and it will be described by a mathematical technique utilizing this feature.

1 is a flowchart illustrating an operation of a distributed machine learning error recovery method according to an embodiment of the present invention, and shows an operation flowchart of a master in a distributed machine learning environment.

Referring to FIG. 1, in the distributed machine learning error recovery method according to an embodiment of the present invention, as shown in FIG. 2, after transmitting the LDPC codes for each of the sub-matrices, if a calculation result value is received from workers, , Among the inspection nodes and the variable nodes, both the variable node corresponding to the erasure and the inspection node connected to the variable node are removed (S110, S120).

After removing all of the variable nodes corresponding to the erasers and the test nodes connected to the variable nodes in step S120, the variable nodes connected to the corresponding test nodes based on the calculation result values stored in the connected variable nodes for each of the remaining test nodes. The first determination value for determining whether there is an error in the field is determined, and an error candidate, that is, an error candidate node, determined to include an error among the variable nodes is determined based on the first determination value determined for each of the inspection nodes. Determine (S130, S140).

If an error candidate is determined by step S140, the error node is determined based on a calculation result value stored in variable nodes (including error candidates) connected to the inspection node including one error candidate among the inspection nodes connected to the error candidate. The calculation result value of the connected error candidate is corrected (S150).

Here, in step S150, the calculation result value of the error candidate may be corrected based on the sum of the calculation result values of all variable nodes connected to the inspection node.

The process of steps S130 to S150 is repeatedly performed a predetermined number of times or when the calculation result values of all error candidates are corrected, the iterative process is terminated, and the calculation result values of the error candidates are based on the calculated result values of the modified variable nodes. To perform erasure decoding (S160).

The present invention will be described in detail with reference to FIGS. 2 to 5.

The present invention relates to a technique for detecting and correcting an incorrect calculation result or malicious information given in distributed machine learning. When an irregular LDPC code is given, the algorithm used in the regular LDPC code can be easily extended and used for convenience of explanation. The present invention will be described on the premise that the proposed code is (d _v , d _c ) -LDPC code.

가 저장되어 있으며, 계산이 끝나지 않은 변수 노드에는 이레이저(erasure)로 표시한다. 변수 노드가 이레이저인 경우 H=[h_mn]_M×N

에서 이레이저인 변수 노드와 그 변수 노드와 연결된 검사 노드를 모두 제거하여 H'=[h'_mn]_M'×N'을 만든다.Initialization steps: In the variable node, the calculation result value calculated by the worker when the calculation is finished

Is stored, and the variable node that has not been calculated is marked with erasure. H = [h _mn ] _{M × N} when the variable node is an _erasure

In this example, H '= [ _h'mn ] _{M' × N 'is created} by removing both the variable node which is an erasure and the check node connected to the variable node.

For example, in FIG. 2, when the variable nodes of B _n-2 and B _{n-1 among} the variable nodes are erasures, that is, by processing straglers as erasures, the variable nodes of the erasure B B _-2 And B _n-1 and B _n-2 and B _n-1, respectively, are removed from the check node (square shape).

,

그리고 복호에 올바른 변수 노드의 집합 B를

φ로 초기화 한다. 각 변수 노드에 대한 판정 값 E_i와 검사 노드에 대한 판정 값 F_i, G_i를 각각 0으로 초기화 한다. If H 'is given, the number of repetitions l is 0, and the value of V _i is calculated by the worker's calculation for each variable node

,

And the correct set of variable nodes B for decoding

Initialize with φ. The determination values E _i for each variable node and the determination values F _i and G _i for the inspection node are initialized to 0, respectively.

Here, the check node determination value F _i is a value for detecting a candidate determined to have an error among the variable nodes, and the check node determination value G _i may be a value for correcting an error. In the initialization step, it can be initialized as shown in <Equation 1> below.

 [Equation 1]

Step 1: If you for all the check nodes c _i satisfies the following <Equation 2>, and displays the F _i to zero, is not satisfied denoted by 1. That is, the inspection node determination value F _i is a value determined according to the sum of the calculation result values (real vector) stored in the variable nodes connected to the corresponding inspection node. When the sum of the calculation result values is 0 vector, all the variable nodes are normal. It can be seen that the calculation result value is stored, and if the sum of the calculation result values is not a vector, at least one variable node among the variable nodes may be regarded as storing an incorrect calculation result value.

[Equation 2]

For example, as shown in FIG. 3, assuming that the variable nodes v ₁ to v ₈ are connected to the check nodes c ₁ to c ₄ , and there is an error in the calculation result value stored in the variable node v ₃ , 4, a variable node v ₃ and connected to the check nodes c ₁ and c ₃ check nodes determination value F _i, as shown in is a 1, a check node is determined the other check node value F _i is zero. Here, a test node having a check node determination value F _i of 1 may be referred to as unsatisfied check nodes. That is, the dissatisfaction check node is checked and the check node determination value F _i of the check node is updated.

Step 2: The variable node v _j that satisfies <Equation 3> below is designated as an error candidate, and E _i = 1. At this time, the variable nodes entering the error candidate may or may not be variable nodes that actually have errors.

That is, the present invention determines the variable node as an error candidate when the check node determination value F _i for each of the check nodes connected to each of the variable nodes is all 1. For example, as illustrated in FIG. 4, by checking the check node determination value F _i for each of the check nodes connected to each of the variable nodes, the variable node v _{3 in} which the check node determination values F _i of the connected check nodes are all 1 is determined. It is determined as an error candidate, and the variable node determination value E _j is 1. That is, an error candidate is determined, and the variable node determination value E _j of the thus determined candidate node is updated.

[Equation 3]

일 수 있다.here,

Can be

| B | = N ', proceed to step 4, otherwise proceed to the next step, step 3. That is, if the error candidate does not exist, the erasure decoding step of step 4 is performed, and when the error candidate exists, a process for correcting the remaining error candidate, that is, the step of step 3 is performed.

Step 3: Calculate the value of G _i for each test node c _i using <Equation 4> below.

 [Equation 4]

That is, Equation (4) calculates the check node decision value G _i as the sum of all of the variable node decision values E _j for each of the check nodes. For example, as illustrated in Figure 5, a check node c ₁ of check node determination value G _i is the first since the E _j values of the variable nodes connected to be 0 and 1, and the check node the check node determines the value of c ₂ G _i becomes 0 because the E _j values of the connected variable nodes are 0, 0, 0, 0, and the check node judgment value of the inspection node c ₃ G _i is the E _j values of the connected variable nodes is 1, 0, 0 Therefore, it becomes 1, and the inspection node determination value G _i of the inspection node c ₄ becomes 0 because the E _j values of the connected variable nodes are 0, 0, and 0.

If the check node judgment value G _i for each check node is calculated, find i with G _i = 1, find the index η not included in B among the variable nodes connected to c _i with respect to _i , and find z _η below. Change as in 5>, and update B using <Equation 6> below.

Here, G _i = 1 means that there is a possibility that only one of the variable nodes connected to c _i may be wrong, and not all variable nodes in the error candidate are in error.

That is, among the check nodes, it is to find the check nodes including only one error candidate in the connected variable nodes, and correct the calculation result value for one error candidate connected to the check node using Equation (2). Mathematics where the sum of the calculated result values of the connected variable nodes is 0 because only one of the variable nodes connected to the corresponding inspection node stores the error value and the calculated result value stored in the remaining variable nodes is a normal result value. Equation 2 can be used to modify the calculation result value of an error candidate storing an error value.

[Equation 5]

[Equation 6]

When the calculation result value of the error candidate is modified through this process, B is updated by adding the index of the variable node whose calculation result value is modified to B.

| B | = N '(all variable nodes are all storing or modifying the result of a normal calculation) or if l is greater than the iteration threshold, proceed to step 4, otherwise update l to l + 1 and Feedback to step 1.

에서 이레이저 복호를 수행한다.Step 4: | B | = N 'means that all variable nodes are valid variable nodes. H = [h _mn ] _{M × N} with a corrected variable node and a straggler node set to erasure in the initialization phase

Eraser decoding is performed.

| B | If <N ', i.e., l is greater than the repetition threshold, the decoding failure is declared.

As described above, the method according to an embodiment of the present invention can detect and correct an incorrect calculation result when workers send an incorrect calculation result value to a master in a distributed machine learning environment using a decoding feature of an LDPC code, and distributed machine learning In an environment, distributed machine learning can be operated by considering both a worker as a straggler and a case in which an incorrect calculation result is transmitted.

In addition, the method according to an embodiment of the present invention can efficiently detect and correct errors that may occur in distributed machine learning using LDPC codes and bipartite graph representations of codes.

6 shows an exemplary diagram for evaluating the performance of the method of the present invention, and shows a graph evaluating the performance of the present invention in the QC LDPC code with (d _v , d _c ) of (6, 12).

Here, the X-axis refers to the number of variable nodes (probability) having an incorrect calculation result value among the variable nodes, and the Y-axis refers to the word error rate (WER).

As shown in FIG. 6, the method according to the present invention can correct all of the incorrect calculation result values when the probability is 0.115, and cannot correct all of the incorrect calculation result values when the probability is 0.12, but for most error candidates It can be seen that the calculation result value can be modified.

7 illustrates a configuration of a distributed machine learning error recovery apparatus according to an embodiment of the present invention, and shows a conceptual configuration of an apparatus performing the method of FIGS. 1 to 6.

Referring to FIG. 7, the apparatus 700 according to an embodiment of the present invention includes a receiver 710, a decision unit 720, a correction unit 730, and a decoding unit 740.

The receiving unit 710 receives a calculation result value performed by workers from workers.

The determination unit 720 determines a first determination value for determining whether there are errors in the variable nodes connected to the corresponding inspection node based on the calculation result values stored in the variable nodes connected to each of the inspection nodes, and the inspection node An error candidate, that is, an error candidate node, which is determined to include an error among the variable nodes is determined based on the first determination value determined for each of them.

At this time, the determination unit 720 may remove the variable node corresponding to the eraser and the inspection node connected to the variable node from among the inspection nodes and the variable nodes, and then determine the first determination value and the error candidate node. have.

When the error candidate is determined, the correction unit 730 checks the error based on the calculation result value stored in the variable nodes (including error candidates) connected to the test node including one of the error candidates connected to the error candidate. Correct the calculation result value of the error candidate connected to the node.

At this time, the correction unit 730 may correct the calculation result value of the error candidate based on the sum of the calculation result values of all variable nodes connected to the inspection node being 0.

Here, the determination unit 720 and the correction unit 730 repeat the process of determining the error candidate and correcting the calculation result value for the error candidate a predetermined number of times after correcting the calculation result value of the error node connected to the inspection node. Can be done.

The decoding unit 740 performs erasure decoding based on the calculation result values of the variable nodes connected to each of the check nodes by correcting the error candidate calculation result value by repeating a predetermined number of times.

Although the description of the apparatus in FIG. 7 is omitted, the apparatus in FIG. 7 may include all of the contents described with reference to FIGS. 1 to 6, and these matters are obvious to those skilled in the art.

The system or device described above may be implemented with hardware components, software components, and / or combinations of hardware components and software components. For example, the systems, devices, and components described in embodiments include, for example, processors, controllers, arithmetic logic units (ALUs), digital signal processors (micro signal processors), microcomputers, field programmable arrays (FPAs). ), A programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions, may be implemented using one or more general purpose computers or special purpose computers. The processing device may run an operating system (OS) and one or more software applications running on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of understanding, a processing device may be described as one being used, but a person having ordinary skill in the art, the processing device may include a plurality of processing elements and / or a plurality of types of processing elements. It can be seen that may include. For example, the processing device may include a plurality of processors or a processor and a controller. In addition, other processing configurations, such as parallel processors, are possible.

The software may include a computer program, code, instruction, or a combination of one or more of these, and configure the processing device to operate as desired, or process independently or collectively You can command the device. Software and / or data may be interpreted by a processing device, or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. , Or may be permanently or temporarily embodied in the transmitted signal wave. The software may be distributed on networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

As described above, although the embodiments have been described by a limited embodiment and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques are performed in a different order than the described method, and / or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form from the described method, or other components Alternatively, even if replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (12)

Receiving calculation result values from workers;
Removing a variable node of an erasure and a check node connected to the variable node of the erasure among variable nodes;
Determining an error candidate among variable nodes based on a calculation result value of variable nodes connected to each of the inspection nodes; And
Correcting the calculated result value of the error node connected to the corresponding inspection node based on the calculated result value of each of the variable nodes connected to the determined error candidate and the connected inspection node.
Distributed machine learning error recovery method comprising a.
According to claim 1,
The determining step
A first judgment value for detecting the error candidate is calculated for each of the inspection nodes based on a calculation result value of variable nodes connected to each of the inspection nodes, and a first judgment calculated for each of the inspection nodes A distributed machine learning error recovery method characterized by determining the error node based on a value.
According to claim 2,
The determining step
A distributed machine learning error recovery method for each of the variable nodes, wherein the variable node is determined as an error candidate when the first determination values of all the inspection nodes connected to the variable node are preset values.
According to claim 2,
The step of correcting
Correcting the calculation result value of the error node connected to the first inspection node so that the sum of the calculation result values of each of the variable nodes connected to the first inspection node including only one of the determined error candidates is a preset value. Distributed machine learning error recovery method characterized in that.
According to claim 4,
After correcting the calculation result value of the error node connected to the first inspection node, repeating the determining step or the correcting step a predetermined number of times.
Distributed machine learning error recovery method further comprising a.
The method of claim 5,
After repeatedly performing the predetermined number of times, performing erasure decoding based on the calculation result values of the variable nodes connected to each of the check nodes.
Distributed machine learning error recovery method further comprising a.
A receiver for receiving calculation result values from workers;
Among variable nodes, the variable node of erasure and the check node connected to the variable node of the eraser are removed, and the variable is based on the calculation result values of the variable nodes connected to each of the inspect nodes. A determination unit for determining error candidates among the nodes; And
Correction to correct the calculation result value of the error node connected to the corresponding inspection node based on the calculated result value of each of the variable nodes connected to the inspection node connected to the determined error candidate.
Distributed machine learning error recovery device comprising a.
The method of claim 7,
The determining unit
A first judgment value for detecting the error candidate is calculated for each of the inspection nodes based on a calculation result value of variable nodes connected to each of the inspection nodes, and a first judgment calculated for each of the inspection nodes A distributed machine learning error recovery apparatus characterized by determining the error node based on a value.
The method of claim 8,
The determining unit
A distributed machine learning error recovery apparatus for each of the variable nodes, wherein the variable node is determined as an error candidate when the first determination values of all the inspection nodes connected to the variable node are preset values.
The method of claim 8,
The correction unit
Correcting the calculation result value of the error node connected to the first inspection node so that the sum of the calculation result values of each of the variable nodes connected to the first inspection node including only one of the determined error candidates is a preset value. Distributed machine learning error recovery device, characterized in that.
The method of claim 10,
The determination unit and the correction unit
Distributed machine learning error characterized by repeating the process of determining the error candidate and correcting the calculation result value for the error candidate a predetermined number of times after correcting the calculation result value of the error node connected to the first inspection node. Recovery device.
The method of claim 11,
After repeating the predetermined number of times, a decoding unit that performs erasure decoding based on the calculation result values of variable nodes connected to each of the check nodes
Distributed machine learning error recovery device further comprising a.

Images