JP2018159992A - Parameter adjustment device, learning system, parameter adjustment method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a parameter adjustment device, a learning system, a parameter adjustment method and a program capable of reducing an adjustment man hour of a hyper parameter.SOLUTION: A parameter adjustment device has an estimation unit, a restriction unit and a determination unit. The estimation unit estimates an estimation function which indicates a relationship between a hyper parameter for specifying operation of machine learning and a learning result, on the basis of the learning result obtained by the machine learning. The restriction unit restricts a value area of the hyper parameter, on the basis of the estimation function estimated by the estimation unit. The determination unit determines the hyper parameter used for the machine learning, from among the hyper parameters contained in the value area restricted by the restriction unit.SELECTED DRAWING: Figure 1

Description

  Embodiments described herein relate generally to a parameter adjustment device, a learning system, a parameter adjustment method, and a program.

  In recent years, research on methods using machine learning has been performed in the fields of image recognition, speech recognition, and language analysis. In particular, a deep learning method using a multilayered neural network has attracted attention. In this deep learning method, learning processing using teacher data or the like is performed in advance to adjust the weight of the neural network connection. When adjusting the connection weight of such a neural network, the values of the hyper parameters (the learning rate, the number of nodes in each layer, the number of layers, etc.) that define the learning operation greatly affect the learning result.

JP 2006-23868 A

  In the learning process as described above, it is necessary to repeat the learning process many times while changing the value of the hyperparameter until a desired learning result is obtained. For example, every time the value of the hyper parameter is changed, it is necessary to confirm the learning result, grasp the relationship between the hyper parameter and the learning result, and then adjust the value of the hyper parameter again. For this reason, it takes time to adjust such hyperparameters.

  In addition, when a general search method based on probability distribution is used to adjust hyperparameters, there is a possibility of repeatedly searching for an unnecessary range that is influenced by the initial value and deviates from the value that should be originally searched. For this reason, a desired learning result may not be obtained even if the hyper parameter is adjusted.

  The problem to be solved by the present invention is to provide a parameter adjustment device, a learning system, a parameter adjustment method, and a program capable of efficiently adjusting hyperparameters.

  The parameter adjustment apparatus according to the embodiment includes an estimation unit, a limitation unit, and a determination unit. The estimation unit estimates an estimation function indicating a relationship between a hyperparameter that defines the operation of the machine learning and the learning result based on a learning result obtained by machine learning. The limiting unit limits the range of the hyperparameter based on the estimation function estimated by the estimation unit. The determining unit determines a hyperparameter used for the machine learning from among the hyperparameters included in the range limited by the limiting unit.

The figure which shows an example of the learning system of embodiment. The flowchart which shows an example of the process of the parameter adjustment apparatus of embodiment. The flowchart which shows an example of the function estimation process of the parameter adjustment apparatus of embodiment. The figure which shows the relationship between the learning result obtained by the learning process in embodiment, and a 1st hyper parameter. The figure which shows the adjustment process of the reference point on the reference function in embodiment, and the reference point in a learning result. The figure which shows the process which estimates the function in which the difference | error of the learning result and calculation result calculated from the reference function in embodiment is the minimum. The figure which shows the several estimation function estimated by the function estimation part in embodiment. The figure which shows the range of the 1st hyperparameter limited by the search range limitation part in embodiment. The flowchart which shows the process of the parameter adjustment apparatus of an Example. The flowchart which shows the function estimation process of the parameter adjustment apparatus of an Example. The figure which shows the relationship between the learning result obtained by the learning process in an Example, and a 1st hyper parameter. The figure which shows the adjustment process of the reference point on the reference function in an Example, and the reference point in a learning result. The figure which shows the process which estimates the function in which the difference | error of the learning result and the calculation result calculated from the reference function in an Example becomes the minimum. The figure which shows the several estimation function estimated by the function estimation part in an Example. The figure which shows the range limited by the search range limitation part in an Example. The figure which shows the learning process result in an Example.

  Hereinafter, a parameter adjustment device, a learning system, a parameter adjustment method, and a program according to embodiments will be described with reference to the drawings.

  FIG. 1 is a diagram illustrating an example of a learning system S according to the embodiment. The learning system S includes, for example, a parameter adjustment device 1 and a learning device 3. The parameter adjustment device 1 and the learning device 3 are connected to each other by a network N. The network N includes, for example, a WAN (Wide Area Network), a LAN (Local Area Network), the Internet, a dedicated line, and the like.

  The parameter adjustment device 1 adjusts hyper parameters in machine learning. That is, when adjusting the hyper parameter of machine learning, the parameter adjustment device 1 estimates the relationship between the hyper parameter and the learning result from the tendency of the learning result (correct answer rate, etc.) obtained by the learning process, and this Limit the hyperparameter range based on the function. For example, the parameter adjustment device 1 estimates the relationship between the hyperparameter and the learning result obtained by the learning device 3 using the hyperparameter as a function, and limits the range of the hyperparameter based on this function. .

  The parameter adjustment device 1 includes, for example, a parameter candidate determination unit 10 (determination unit), a task transmission unit 12, a function estimation unit 14 (estimation unit), a search range limitation unit 16 (limitation unit), and a storage unit 18. Is provided. Some or all of the functional units of the parameter adjustment apparatus 1 may be realized by a processor executing a program (software). In this case, the parameter adjustment apparatus 1 may be realized by installing the above program in a computer device in advance. Alternatively, the above program stored in a storage medium such as a CD-ROM or the above program distributed via a network may be appropriately installed in a computer device.

  The parameter candidate determination unit 10 determines a candidate for a combination of a hyperparameter type and a hyperparameter value to be adjusted. The parameter candidate determination unit 10 uses a random search method based on a uniform distribution, a Bayesian search method based on a probability distribution, and the like to determine combination candidates.

  The task transmission unit 12 transmits a task indicating a learning process using the candidate determined by the parameter candidate determination unit 10 to the learning device 3. The learning device 3 performs a learning process based on this task.

  The function estimation unit 14 estimates a function (hereinafter referred to as “estimation function”) indicating the relationship between the hyperparameter and the learning result based on the tendency of the learning result received from the learning device 3.

  Based on the estimation function estimated by the function estimation unit 14, the search range limiting unit 16 predicts the tendency of the learning result including the unlearned range where the learning result is not obtained, and determines the optimal learning result from the predicted tendency. Limits the range of hyperparameters that are expected to be obtained.

  The storage unit 18 stores a hyper parameter search range used in machine learning in advance. In addition, the storage unit 18 stores the range of each hyper parameter limited by the search range limitation unit 16. The storage unit 18 is realized by a ROM (Read Only Memory), a RAM (Random Access Memory), a HDD (Hard Disk Drive), a flash memory, or the like.

  The learning device 3 performs determination processing such as image recognition, voice recognition, and language analysis. The learning device 3 performs a learning process based on the hyperparameter received from the parameter adjustment device 1. The learning device 3 uses, for example, a deep learning method using a multilayered neural network. In this learning process, the configuration of the neural network, the connection weight, and the like are adjusted. The learning device 3 inputs a learning result (correct answer rate, etc.) in the learning process to the parameter adjustment device 1.

  Next, the operation of the parameter adjustment device 1 will be described. FIG. 2 is a flowchart illustrating an example of processing of the parameter adjustment apparatus 1.

  First, the parameter candidate determination unit 10 determines a hyper parameter to be adjusted among hyper parameters such as a learning rate, the number of nodes in each layer, and the number of layers based on a user instruction input from an input unit (not shown). A parameter is selected (step S101). In the following, an example will be described in which two hyper parameters (first hyper parameter A and second hyper parameter B) related to the learning process are to be adjusted.

Next, the parameter candidate determination unit 10 determines a plurality of candidates for the value of one of the two hyper parameters (the first hyper parameter A). That is, the parameter candidate determination unit 10 reads the search range of the first hyperparameter A from the storage unit 18 and determines a plurality of value candidates for the first hyperparameter A within the search range. As the search range of the first hyperparameter A, for example, a large range such as 10 ⁿ (n = 1, 2,...) May be set. Further, the parameter candidate determination unit 10 sets the other hyper parameter (second hyper parameter) as a fixed value.

  Next, the task transmission unit 12 learns the parameter set determined by the parameter candidate determination unit 10 (a combination of each of the first hyperparameter A having a plurality of values and the second hyperparameter B having a fixed value). 3 to cause the learning device 3 to perform a learning process (step S105). In addition, the task transmission unit 12 may transmit a hyper parameter that defines the number of repetitions of the learning process performed in the learning device 3 to the learning device 3.

  Next, the function estimation unit 14 estimates an estimation function indicating the relationship between the first hyperparameter A and the learning result based on the learning result received from the learning device 3 (step S107). FIG. 3 is a flowchart illustrating an example of the function estimation process of the parameter adjustment apparatus 1.

  In the function estimation process, first, the function estimation unit 14 sets the reference point so that the reference point in the learning result received from the learning device 3 is equal to the reference point of the reference function f (A) defined in advance. The function f (A) is multiplied by α (constant) (step S201). As the reference function f (A), a plurality of reference functions f (A) may be defined in advance. In this case, a function indicating a tendency close to the learning result received from the learning device 3 may be selected from the plurality of reference functions f (A). FIG. 4 is a diagram showing the relationship between the learning result (correct answer rate) obtained by the learning process by the learning device 3 and the first hyperparameter A. FIG. 5 is a diagram illustrating an adjustment process between the reference point D1 on the reference function f (A) and the reference point D2 in the learning result. As shown in FIG. 5, the reference function f (A) is multiplied by α so that the correct answer rate values of the reference point D1 on the reference function f (A) and the reference point D2 in the learning result are equal.

  Next, the function estimation unit 14 determines an error between the learning result received from the learning device 3 and the calculation result calculated from the function α * f (A) obtained by multiplying the reference function f (A) by α. The function having the smallest (difference) is estimated as the first estimation function F1 (step S203). For example, as illustrated in FIG. 6, the function estimation unit 14 calculates a function that minimizes the sum of errors between the learning result received from the learning device 3 and the calculation result calculated from the reference function α * f (A). By determining, the first estimation function F1 is estimated.

  Next, the function estimation unit 14 estimates the second estimation function F2 corresponding to the minimum value of the second hyperparameter B using the reference function f (A) (step S205). For example, the function estimation unit 14 transmits a combination of each of the first hyperparameter A having a plurality of values determined by the parameter candidate determination unit 10 and the minimum value of the second hyperparameter B to the learning device 3. The second estimation function is a function that minimizes the error between the learning result obtained by performing the learning process and the calculation result calculated from the function obtained by multiplying the reference function f (A) by a constant. Estimated as F2.

  Next, the function estimation unit 14 estimates the third estimation function F3 corresponding to the maximum value of the second hyperparameter B using the reference function f (A) (step S207). For example, the function estimation unit 14 transmits a combination of each of the first hyperparameter A having a plurality of values determined by the parameter candidate determination unit 10 and the maximum value of the second hyperparameter B to the learning device 3. The third estimation function is a function that minimizes the error between the learning result obtained by performing the learning process and the calculation result calculated from the function obtained by multiplying the reference function f (A) by a constant. Estimated as F3.

  Next, the function estimation unit 14 estimates an estimation function in a range in which a learning result is not obtained (unlearned range) based on the first estimation function F1, the second estimation function F2, and the third estimation function F3. (Step S209). For example, the function estimation unit 14 estimates the change of the other parameter C associated with the second hyperparameter B in each estimation function, thereby the second hyperparameter B set as a fixed value, the second An estimation function between the minimum value of the hyperparameter B is estimated. The function estimation unit 14 also estimates an estimation function between the value of the second hyperparameter B set as a fixed value and the maximum value of the second hyperparameter B.

  FIG. 7 is a diagram illustrating a plurality of estimation functions estimated by the function estimation unit 14. In the example illustrated in FIG. 7, the first estimation function F1, the second estimation function F2, and the third estimation function F3 are estimated by the function estimation unit 14, and then associated with the second hyperparameter B in each estimation function. By estimating the change in parameter C, the estimation function (dotted line) in the unlearned range other than the estimation functions F1, F2, and F3 is estimated. Thus, the process of this flowchart relating to function estimation ends.

  Next, the search range limiting unit 16 limits the range of the first hyperparameter A in each value of the second hyperparameter B based on the plurality of estimation functions estimated by the function estimation unit 14 (step S111). When selecting the value of the second hyperparameter B, a random method based on a uniform distribution, a metropolis method, or the like may be used. Further, when the range of the first hyperparameter A is limited, a predetermined range based on the peak value of the learning result in each estimation function may be limited as the range. The search range limiting unit 16 stores the limited value range in the storage unit 18.

  FIG. 8 is a diagram illustrating a range of the first hyperparameter A limited by the search range limiting unit 16. In the example illustrated in FIG. 8, a predetermined range σ based on the learning result peak P in the estimation function estimated by the function estimation unit 14 is limited as a range.

  Next, the parameter candidate determining unit 10 predicts that an optimal learning result is obtained from the first hyperparameter A included in the range limited by the search range limiting unit 16 stored in the storage unit 18. The value of the first hyperparameter A to be selected is selected, and the parameter set of the selected first hyperparameter A and second hyperparameter B is determined (step S113).

  Next, the task transmission unit 12 transmits the parameter set used for machine learning determined by the parameter candidate determination unit 10 to the learning device 3, and causes the learning device 3 to perform a learning process (step S115). Thus, the process of this flowchart is completed.

  According to the embodiment described above, the number of searches for an unnecessary range is reduced by estimating the relationship between each parameter and the learning result from a function of the actual learning result and learning while limiting the range. It is possible to reduce the man-hours required for adjusting the hyper parameters.

  Next, an embodiment relating to a hyper parameter adjustment method using Momentum SGC will be described. FIG. 9 is a flowchart illustrating processing of the parameter adjustment device 1 in the embodiment.

  First, the parameter candidate determination unit 10 determines a hyper parameter to be adjusted among hyper parameters such as a learning rate, the number of nodes in each layer, and the number of layers based on an operator instruction input from an input unit (not shown). A parameter is selected (step S301). Here, the parameter candidate determination unit 10 selects “learning rate” as the first hyperparameter A to be adjusted, and selects “momentum” as the second hyperparameter B.

Next, the parameter candidate determination unit 10 determines “learning rate = 10 ⁻ⁿ (n = 1, 2,..., 5)” as the first hyperparameter A (learning rate), and the second hyperparameter B ( B _default ) (momentum) is determined as “momentum = 0.90” (step S303).

  Next, the task transmission unit 12 transmits the parameter set determined by the parameter candidate determination unit 10 (a combination of each of the five first hyperparameters A and the fixed second hyperparameter B) to the learning device 3. Then, the learning apparatus 3 is caused to perform learning processing (step S305). FIG. 11 is a diagram illustrating the relationship between the learning result obtained by the learning process and the first hyperparameter A.

  Next, the function estimation unit 14 estimates an estimation function indicating the relationship between the first hyperparameter A and the learning result based on the learning result received from the learning device 3 (step S307). FIG. 10 is a flowchart showing the function estimation process of the parameter adjustment apparatus 1.

In the function estimation process, first, the function estimation unit 14 sets the reference point so that the reference point in the learning result received from the learning device 3 and the reference point on the predefined reference function f (A) are equal. The function f (A) is multiplied by α (constant) (step S401). Here, it is assumed that the Poisson distribution of the following equation (1) is defined as the reference function f (A).

K in the above equation (1) satisfies the relationship of n = k + 1 with n in the learning rate (10 ⁻ⁿ ). That is, k in the equation (1) is a variable associated with the first hyperparameter A. In this equation (1), λ is a variable (parameter C) associated with the second hyperparameter B.

  As shown in FIG. 12, the reference function f (A) is multiplied by α so that the correct answer rate values of the reference point D1 on the reference function f (A) and the reference point D2 in the learning result are equal. For example, the function estimation unit 14 sets the reference function so that the value of the distribution when k = 1 (learning rate = 0.01) is equal to the value of the learning result (correct answer rate) received from the learning device 3. Multiply f (A) by α.

Next, the function estimation unit 14 determines an error between the learning result received from the learning device 3 and the calculation result calculated from the function α * f (A) obtained by multiplying the reference function f (A) by α. A function having the minimum (difference) is estimated as the first estimation function F1 (step S403). For example, the function estimation unit 14 estimates the value of λ that minimizes the sum of square errors. As illustrated in FIG. 13, the function estimation unit 14 estimates a function having λ ₀ = 2 in the equation (1) as the first estimation function F1.

Next, the function estimation unit 14 estimates the second estimation function F2 corresponding to the minimum value (B _min ) of the second hyperparameter B using the reference function f (A) (step S405). For example, the function estimation unit 14 gives the learning device 3 a combination of each of the five first hyperparameters A determined by the parameter candidate determination unit 10 and the minimum value (0.01) of the second hyperparameter B. A second function that minimizes the error between the learning result obtained by transmitting and performing the learning process and the calculation result calculated from the function obtained by multiplying the reference function f (A) by a constant is the second function. Estimation is performed as an estimation function F2.

Next, the function estimation unit 14 estimates the third estimation function F3 corresponding to the maximum value (B _max ) of the second hyperparameter B using the reference function f (A) (step S407). For example, the function estimation unit 14 gives a combination of each of the five first hyperparameters A determined by the parameter candidate determination unit 10 and the maximum value (0.99) of the second hyperparameter B to the learning device 3. A function that minimizes an error between a learning result obtained by transmitting and performing a learning process and a calculation result calculated from a function obtained by multiplying the reference function f (A) by a constant is a third function. Estimation is performed as an estimation function F3.

Next, the function estimation unit 14 estimates an estimation function in a range in which a learning result is not obtained (unlearned range) based on the first estimation function F1, the second estimation function F2, and the third estimation function F3. (Step S409). As shown in FIG. 14, after estimating the first estimation function F1, the second estimation function F2, and the third estimation function F3, the function estimation unit 14 performs the first estimation function in the range of 0.01 <momentum <0.99. 2 An approximate expression representing the relationship between the hyperparameter B and the parameter λ is obtained. For example, the function estimation unit 14 obtains an approximate expression (4) that represents the relationship between the following expressions (2) and (3). In formula (4), a and b are constants.

Here, since λ ₀ = B _default and ΔB = B−B _default (B is given by the probability distribution at the time of learning), the approximate expression (4) is expressed as Expression (5).

In this embodiment, the function estimation unit 14 obtains an approximate expression represented by the following expression (6) (b in the above expression (5) is 0). In this approximate expression (6), β is a constant. The function estimation unit 14 obtains the value of λ at each momentum value in increments of 0.01 using this approximate expression (6) (except momentum = 0.9).

Thus, the estimation function in the unlearned range is estimated by estimating the change of the parameter λ associated with the second hyperparameter B in the first estimation function F1, the second estimation function F2, and the third estimation function F3. Is done. Thus, the process of this flowchart relating to function estimation ends. Note that the approximate expression obtained by the function estimation unit 14 need not be a linear function expression, but may be an expression such as a quadratic function. In addition, when the approximate expression is defined in advance, it corresponds to the second estimation function F2 corresponding to the minimum value (B _min ) of the second hyperparameter B and the maximum value (B _max ) of the second hyperparameter B. The process of estimating the third estimation function F3 may be omitted.

  Next, the search range limiting unit 16 limits the range of the first hyperparameter A based on the plurality of estimation functions estimated by the function estimation unit 14 (step S311). FIG. 15 is a diagram illustrating a plurality of estimation functions estimated by the function estimation unit 14 and the range of values limited by the search range limitation unit 16. In the example shown in FIG. 15, in the estimation function estimated for each of the four second hyperparameters B (0.09, 0.62, 0.90, 0.99), correct answer rate peaks P1 to P4 are used. Is limited as the range of the first hyperparameter A. The search range limiting unit 16 may limit a predetermined range based on each of the peaks P1 to P4 as a range.

  Next, the parameter candidate determination unit 10 selects the first hyperparameter A used for machine learning from the first hyperparameters A included in the range limited by the search range limitation unit 16, and selects the selected first hyperparameter. A parameter set of the parameter A and the second hyper parameter B is determined (step S313).

  Next, the task transmission unit 12 transmits the parameter set used for machine learning determined by the parameter candidate determination unit 10 to the learning device 3, and causes the learning device 3 to perform a learning process (step S315). Thus, the process of this flowchart is completed.

  FIG. 16 is a diagram illustrating a processing result when the learning process is performed using the hyper parameter adjusted by the parameter adjusting device 1 according to the embodiment. The average improvement rate of the correct answer rate indicates the difference between the correct answer rate of the reference argument (lr = 0.01, momentum = 0.9) and the maximum value of the correct answer rate obtained by the adjustment. The average search number indicates the average value of the search number until a value equal to or higher than the correct answer rate obtained in the example is obtained in the Bayesian search. Since the maximum number of searches in Bayesian search is 40, the value is calculated as 40 if it cannot be obtained within that number.

  As shown in FIG. 16, when the learning process is performed using the hyperparameter adjusted by the parameter adjustment device 1 in the embodiment as compared with the case where the conventional Bayesian search is used, the average improvement width of the same degree is obtained. By value, the average number of searches can be reduced to about half. Further, according to the result of the search number of 15 times, the value of the average improvement width can be improved by the present embodiment. For example, when compared under the condition where the number of classes is “30” and the number of searches is “15”, the average improvement width when the conventional Bayesian search is used is “1.81”, whereas In this example, it is “2.52”, which indicates that the value of the average improvement width is improved. Further, as the number of classes increases, the average improvement width can be further improved in the present embodiment.

  According to the embodiment described above, the number of searches for an unnecessary range is reduced by estimating the relationship between each parameter and the learning result from a function of the actual learning result and learning while limiting the range. It is possible to reduce the man-hours required for adjusting the hyper parameters.

  In the above-described embodiment, an example in which two parameters (first hyperparameter A and second hyperparameter B) are used as parameters to be adjusted has been described. However, the present embodiment is used to adjust three or more parameters. You may apply.

  Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 1 ... Parameter adjustment apparatus, 3 ... Learning apparatus, 10 ... Parameter candidate determination part, 12 ... Task transmission part, 14 ... Function estimation part, 16 ... Search range limitation part, 18 ... Memory | storage part, S ... Learning system

Claims (9)

Based on a learning result obtained by machine learning, an estimation unit that estimates an estimation function indicating a relationship between a hyperparameter that defines the operation of the machine learning and the learning result;
Based on the estimation function estimated by the estimation unit, a limiting unit that limits the range of the hyperparameter;
A parameter adjustment device comprising: a determination unit that determines a hyperparameter used for the machine learning from among the hyperparameters included in the range limited by the limitation unit. The estimation unit estimates the estimation function including a range in which a learning result of the machine learning is not obtained;
The parameter adjustment device according to claim 1. The estimation unit uses, as the estimation function, a reference function that minimizes a difference between the learning result obtained by the machine learning and a calculation result calculated based on the reference function among a plurality of reference functions. presume,
The parameter adjustment apparatus according to claim 1 or 2. The determination unit determines a hyperparameter used for the machine learning based on a search method using a uniform distribution.
The parameter adjustment apparatus as described in any one of Claim 1 to 3. The estimation unit estimates an estimation function indicating a relationship between a hyperparameter used for adjusting a configuration of a neural network and a connection weight and the learning result.
The parameter adjustment device according to any one of claims 1 to 4. The estimation unit estimates, as the estimation function, a function that minimizes a difference between the learning result obtained by the machine learning and a calculation result calculated based on a Poisson distribution function;
The parameter adjustment device according to any one of claims 1 to 5. A parameter adjusting device according to any one of claims 1 to 6;
A learning system comprising: a learning device that performs learning processing using the hyperparameter determined by the parameter adjustment device. Based on a learning result obtained by machine learning, an estimation function indicating a relationship between the learning result and a hyperparameter that defines the operation of the machine learning is estimated,
Based on the estimated estimation function, a range of the hyperparameter is limited,
Determining hyperparameters used for the machine learning from among hyperparameters included in the limited range;
Parameter adjustment method. On the computer,
Based on a learning result obtained by machine learning, an estimation function indicating a relationship between a hyperparameter that defines the operation of the machine learning and the learning result is estimated,
Based on the estimated estimation function, the range of the hyperparameter is limited,
From among the hyperparameters included in the limited range, the hyperparameter used for the machine learning is determined.
program.

Images