JP2023124376A - Information processing apparatus, information processing method, and program - Google Patents

Abstract

To establish a high-accuracy neural network.SOLUTION: A method includes the steps of: outputting input data and a first inference result of the input data on the basis of a first neural network; outputting a first evaluation result based on the first inference result; updating a first weight parameter of the first neural network on the basis of the first evaluation result; outputting a second inference result of the input data on the basis of the input data, the first neural network set with the updated first weight parameter, a second neural network, and a weight coefficient of the second neural network; outputting a second evaluation result based on the second inference result; updating the weight coefficient on the basis of the second evaluation result; and updating a second weight parameter on the basis of the updated first weight parameter, the second weight parameter of the second neural network, and the updated weight coefficient.SELECTED DRAWING: Figure 1

Description

The present invention relates to an information processing device, an information processing method, and a program.

A neural network (hereinafter also referred to as "NN") has high performance in problems such as image recognition. In the field utilizing the NN, learning data is used to learn the NN, and evaluation data is used to verify the performance of the NN. However, the parameters of the NN undergo drastic changes during learning due to their dependence on the training data, and drastic changes in the parameters may lead to over-learning. Over-learning is excessive adaptation to learning data, and refers to a phenomenon in which accuracy with respect to evaluation data deteriorates.

In recent years, in order to solve this problem, a method for improving robustness by applying an exponentially moving average (EMA) to the NN parameters is often adopted when storing the NN. EMA is a technique in which a verification NN is prepared separately from a learning NN, and a weighted sum of the parameters of the learning NN and the verification NN is obtained. This method improves the generalization performance, and high-precision identification can be expected even in the evaluation data. The processing of EMA is shown as Equation (1).

The method described in Non-Patent Document 1 is one of the methods of semi-supervised learning using EMA, especially in image recognition, and the NN saved by EMA is more effective in verification data than the learned NN. This indicates that the accuracy is high. Here, semi-supervised learning is a method of learning in a state in which teacher labels exist only in part of the huge amount of input data and teacher labels required for NN learning. Work load can be reduced.

The method described in Non-Patent Document 2 is one of the learning methods using EMA for learning of a generative adversarial network (GAN). This indicates that learning progresses more stably. Here, GAN is a technique for learning the characteristics of data input to the NN, and can generate non-existent but high-quality data and convert data in line with the characteristics of the input data.

Zhaowei Cai, 5 others, "Exponential Moving Average Normalization for Self-supervised and Semi-supervised Learning", [online], [searched on February 9, 2020], Internet <https://arxiv.org/abs/ 2101.08482> Yasin Yazici, 5 others, "THE UNUSUAL EFFECTIVENESS OF AVERAGING IN GAN TRAINING", [online], [searched on February 9, 2020], Internet <https://arxiv.org/abs/1806.04498>

In the method described in Non-Patent Document 1 and the method described in Non-Patent Document 2, the coefficient α in Equation (1) is always constant, and is a parameter that is manually set at the start of learning. However, it is conceivable that the appropriate coefficient α differs depending on the NN to be learned or the data set to be used. Therefore, the user needs to try learning the NN a plurality of times while changing the coefficient α, or analyze the NN after learning, which increases the burden on the user.

Therefore, the present invention has been proposed to solve these problems, and it is desirable to provide a technique that enables construction of a highly accurate neural network while reducing the burden on the user. be

In order to solve the above problem, according to one aspect of the present invention, a first input data and a first neural network output a first inference result of the first input data. 1 calculation unit, a first evaluation unit that outputs a first evaluation result based on the first inference result, and a first weight parameter of the first neural network based on the first evaluation result a first update unit that updates the second input data; the first neural network in which the updated first weight parameter is set; a second neural network; and the second neural network a second calculation unit for outputting a second inference result of the second input data based on the weighting coefficient of the network; and a second calculation unit for outputting a second evaluation result based on the second inference result. an evaluation unit, a weighting factor updating unit that updates the weighting factor based on the second evaluation result, the updated first weighting parameter, and a second weighting parameter of the second neural network; and a second updating unit that updates the second weighting parameter based on the updated weighting factor.

The second computing unit inputs the second input data to the first layer of each of the first neural network and the second neural network to which the updated first weight parameter is set. can be entered.

The second computing unit performs the following operations on the second and subsequent layers of each of the first neural network and the second neural network to which the updated first weight parameter is set: A weighted sum of the weighting factors for the output may be input.

The second evaluation unit may calculate a loss based on the second evaluation result and the teacher label, and output the second evaluation result based on the loss.

The weighting factor updating unit may update the weighting factor by error backpropagation based on the second evaluation result.

The second updating unit calculates the second weighting parameter by a weighted sum of the updated first weighting parameter and the second weighting parameter of the second neural network using the updated weighting factor. Weight parameters may be updated.

The weighting factor may be provided for each layer included in the first neural network and the second neural network.

The first input data and the second input data may be the same data.

The first input data and the second input data may be different data.

According to another aspect of the present invention, outputting a first inference result of the first input data based on the first input data and a first neural network; outputting a first evaluation result based on the inference result of; updating a first weight parameter of the first neural network based on the first evaluation result; and second input data; Based on the first neural network set with the updated first weighting parameter, the second neural network, and the weighting factor of the second neural network, the second input data outputting an inference result of 2; outputting a second evaluation result based on the second inference result; updating the weighting factor based on the second evaluation result; updating the second weighting parameter based on the subsequent first weighting parameter, the second weighting parameter of the second neural network, and the updated weighting factor. A processing method is provided.

According to another aspect of the present invention, a computer is provided with a first input data and a first neural network for outputting a first inference result of the first input data. a calculation unit, a first evaluation unit that outputs a first evaluation result based on the first inference result, and a first weight parameter of the first neural network that is updated based on the first evaluation result. a first update unit that performs the above, second input data, the first neural network in which the updated first weight parameter is set, a second neural network, and the second neural network a second calculation unit that outputs a second inference result of the second input data based on the weighting factor; and a second evaluation unit that outputs a second evaluation result based on the second inference result. a weighting factor updating unit that updates the weighting factor based on the second evaluation result; the updated first weighting parameter; the second weighting parameter of the second neural network; and a second updating unit that updates the second weighting parameter based on the weighting factor afterward.

As described above, according to the present invention, a technique is provided that enables construction of a highly accurate neural network while reducing the burden on the user.

1 is a diagram showing a functional configuration example of a learning device according to a first embodiment of the present invention; FIG. It is a figure for demonstrating each process of the calculating part and double calculating part which concern on the same embodiment. 5 is a flow chart showing an example of operation in a learning stage performed by the learning device according to the embodiment; It is a figure for demonstrating the process of the double operation part which concerns on the same embodiment. 1 is a diagram showing a hardware configuration of an information processing device as an example of a learning device according to a first embodiment of the present invention; FIG.

Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. In the present specification and drawings, constituent elements having substantially the same functional configuration are denoted by the same reference numerals, thereby omitting redundant description.

In addition, in this specification and drawings, a plurality of components having substantially the same functional configuration may be distinguished by attaching different numerals after the same reference numerals. However, when there is no particular need to distinguish between a plurality of constituent elements having substantially the same functional configuration, only the same reference numerals are used. Also, similar components in different embodiments may be distinguished by attaching different alphabets after the same reference numerals. However, when there is no particular need to distinguish between similar components of different embodiments, only the same reference numerals are used.

(0. Outline of embodiment)
An outline of an embodiment of the present invention will be described. In the embodiment of the present invention, an information processing device (hereinafter also referred to as a "learning device") that performs neural network learning will be described. In the learning device, learning of the neural network is performed based on the learning data (learning stage). After that, the identification device outputs the identification result based on the trained neural network and identification data (test data) (inference stage).

In the embodiments of the present invention, it is mainly assumed that the learning device and the identification device are realized by the same computer. However, the learning device and the identification device may be implemented by separate computers. In such a case, the trained neural network generated by the learning device is provided to the identification device. For example, the trained neural network may be provided from the learning device to the identification device via a recording medium or via communication.

The “learning stage” performed in the learning device will be described below. Below, a neural network is also written as "NN."

(1. First embodiment)
First, a first embodiment of the present invention will be described.

(Structure of learning device)
A configuration example of a learning device according to the first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing a functional configuration example of a learning device 10 according to the first embodiment of the present invention. As shown in FIG. 1, the learning device 10 according to the first embodiment of the present invention includes a learning data set 101, an input unit 102, a learning NN 110, an evaluation unit 121, an update unit 122, It includes a verification NN 130 , an EMA evaluation unit 141 , an EMA update unit 142 and a storage unit 143 .

The learning NN 110 corresponds to an example of a first neural network and is configured including weight parameters 111 . Weight parameter 111 may correspond to an example of a first weight parameter. The learning device 10 also includes a computing unit 112 that performs computation using the learning NN 110 in which the weighting parameter 111 is set. Verification NN 130 corresponds to an example of a second neural network, and includes EMA weighting parameters 131 and EMA coefficients 132 . EMA weight parameter 131 may be an example of a second weight parameter. The learning device 10 includes a double calculator 133 that performs calculations using EMA weighting parameters 131 and EMA coefficients 132 .

The input unit 102, the calculation unit 112, the evaluation unit 121, the update unit 122, the double calculation unit 133, the EMA evaluation unit 141, the EMA update unit 142, the storage unit 143, and the like are CPU (Central Processing Unit) or GPU (Graphics Processing Unit). ), and a program stored in a ROM (Read Only Memory) is loaded into a RAM by the arithmetic device and executed, thereby realizing its function.

At this time, a computer-readable recording medium recording the program may also be provided. Alternatively, these blocks may be composed of dedicated hardware, or may be composed of a combination of multiple pieces of hardware. Data necessary for calculation by the calculation device are appropriately stored in a storage unit (not shown).

The learning data set 101, weight parameters 111, EMA weight parameters 131 and EMA coefficients 132 are stored in a storage unit (not shown). The storage unit may be composed of a memory such as a RAM (Random Access Memory), a hard disk drive, or a flash memory.

In the initial state, the weight parameter 111, the EMA weight parameter 131, and the EMA coefficient 132 are each set to an initial value. For example, the initial values set for the weight parameter 111, the EMA weight parameter 131, and the EMA coefficient 132 may be random values, but may be any values. For example, the initial values set for the weight parameter 111, the EMA weight parameter 131, and the EMA coefficient 132 may be learned values obtained in advance through learning.

(Learning data set 101)
The learning data set 101 includes a plurality of pieces of learning data (hereinafter also referred to as “input data”). A teacher label may be associated with each of the plurality of pieces of input data. Note that the embodiment of the present invention mainly assumes that the input data is image data (especially still image data). However, the type of input data is not particularly limited, and data other than still image data can be used as input data. For example, the input data may be moving image data including a plurality of frames, time-series data, or audio data.

(Input unit 102)
The input unit 102 sequentially acquires input data from the learning data set 101, creates mini-batches based on the acquired input data, and distributes the created mini-batches to the computing unit 112 of the learning NN 110 and the verification NN 130. It outputs to each of the calculation units 133 . The size of the mini-batch is not particularly limited. Also, here, the learning NN 110 and the verification NN 130
It is mainly assumed that the input data input to each are the same data. However, the input data input to the learning NN 110 and the verification NN 130 may be different data.

Each of the arithmetic unit 112 and the double arithmetic unit 133 is configured by connecting computation graphs constructed by neurons in order of processing, and can be regarded as one neural network as a whole. More specifically, each of the computing unit 112 and the double computing unit 133 may mainly include a convolutional layer, a pooling layer and an activation function. In the following, it is mainly assumed that a two-dimensional convolutional layer is used as the convolutional layer, but a three-dimensional convolutional layer may also be used.

(Calculation unit 112)
The calculation unit 112 obtains an inference result based on the input data included in the mini-batch output from the input unit 102 and the learning NN 110 . The computing unit 112 can correspond to an example of a first computing unit. Also, the inference result obtained by the calculation unit 112 can correspond to an example of the first inference result. More specifically, the calculation unit 112 obtains the data output from the learning NN 110 as an inference result based on inputting the input data included in the mini-batch to the learning NN 110 to which the weight parameter 111 is set.

Note that the format of the inference result output from the calculation unit 112 is not particularly limited. However, it is preferable that the format of the inference result output from the calculation unit 112 is set together with the format of the teacher label. For example, if the teacher label indicates the class of the classification problem and is a one-hot vector having the length of the number of classes, the format of the inference result output from the calculation unit 112 is also a vector having the length of the number of classes. can be At this time, the inference result output from the calculation unit 112 may include a value for each class (hereinafter also referred to as “inference value”).

As an example, when the calculation unit 112 adjusts the sum of the inference values of all classes to 1, the inference value corresponding to each class can correspond to the probability corresponding to each class. However, the sum of the inference values of all classes does not have to be adjusted to be 1 by the calculation unit 112 . In any case, the inference value output from the calculation unit 112 can be a larger value as the probability of the class is higher.

Here, the processing of the calculation unit 112 will be described with reference to FIG.

FIG. 2 is a diagram for explaining processing of each of the arithmetic unit 112 and the double arithmetic unit 133 according to the first embodiment of the present invention. Referring to FIG. 2, a configuration example of the calculation unit 112 is shown. In the example shown in FIG. 2, among the plurality of layers of the learning NN 110 used by the calculation unit 112, the i-th layer (where i=1, 2, . The convolutional layer is denoted as "f _i ". A weight parameter corresponding to the i-th layer among the weight parameters 111 is set for the convolution layer f _i .

Also, the input to the convolution layer f _i is expressed as x _i and the output from the convolution layer f _i is expressed as y _i . However, the i-th processing layer of the learning NN 110 may not be a convolutional layer. The input _x1 to the first layer is the input data output from the input section 102 . The input x _i (i≧2) to the second and subsequent layers is the output y _i−1 from the immediately preceding layer. The calculation unit 112 outputs the output y _n corresponding to the final layer of the learning NN 110 to the evaluation unit 121 as an inference result.

Returning to FIG. 1, the description continues.

(Evaluation unit 121)
The evaluation unit 121 obtains an evaluation result based on the inference result for each mini-batch. The evaluator 121 may correspond to an example of a first evaluator. More specifically, the evaluation unit 121 calculates the loss based on the inference result and target for each mini-batch, and obtains the evaluation result based on the loss for each mini-batch. The evaluation unit 121 outputs the evaluation result to the updating unit 122 .

Here, the goal is not limited to a specific goal, and a goal similar to that used in general neural networks may be used. For example, when supervised learning is performed, the goal may be the supervised label corresponding to the input data.

The loss function used to calculate the loss is not limited to a specific function, and loss functions similar to loss functions used in general neural networks may be used. For example, the loss function may be the mean squared error based on the difference between the teacher label and the inference result, or the cross-entropy error based on the difference between the teacher label and the inference result.

(Update unit 122)
The update unit 122 updates the weight parameter 111 based on the evaluation result output from the evaluation unit 121 . This allows the weight parameter 111 to be trained so that the inference result approaches the teacher label. The updating unit 122 can correspond to an example of a first updating unit. More specifically, the updating unit 122 may update the weight parameter 111 by error back propagation (back propagation) based on the evaluation result.

Note that the update unit 122 determines whether or not the learning termination condition of the learning NN 110 is satisfied each time the update of the weight parameter 111 is completed. When it is determined that the learning end condition is not satisfied, the next input data is acquired by the input unit 102, and the calculation unit 112, the evaluation unit 121, and the update unit 122 each perform processing based on the next input data. Each process is executed again. On the other hand, when it is determined that the end condition of learning of the learning NN 110 is satisfied, the learning is ended.

The condition for ending the learning of the learning NN 110 is not particularly limited, and any condition may be used as long as it indicates that the learning of the learning NN 110 has been completed to some extent.

Specifically, the learning end condition of the learning NN 110 may include a condition that the loss is less than the threshold. Alternatively, the learning termination condition of the learning NN 110 may include a condition that the change in the loss is smaller than a threshold (a condition that the loss has converged). Alternatively, the learning end condition of the learning NN 110 may include a condition that the weight parameter 111 has been updated a predetermined number of times. Alternatively, when the accuracy of the learning NN 110 (for example, accuracy rate) is calculated, the conditions for ending the learning of the learning NN 110 may include a condition that the accuracy exceeds a predetermined percentage (for example, 90%). .

(Double operation unit 133)
The double computing unit 133 receives the input data included in the mini-batch output from the input unit 102, the learning NN 110 set with the weighting parameter 111 updated by the updating unit 122, the verification NN 130, and the EMA coefficient 132. An inference result (hereinafter also referred to as an "EMA inference result") is obtained based on. An EMA inference result may correspond to an example of a second inference result. The double computing unit 133 can correspond to an example of a second computing unit. The EMA coefficient 132 is a scalar coefficient and corresponds to the weighting coefficient of the verification NN 130 . Since the EMA coefficient 132 changes at each time t, the EMA coefficient 132 at time t is also expressed as a weighting coefficient α _t below.

Note that the time t indicates the number of updates corresponding to the weight parameter 111, the EMA weight parameter 131, and the EMA coefficient 132, respectively. Here, as an example, assume that the initial value of time t is 0, and time t corresponding to the weighting parameter or coefficient is increased by 1 each time the weighting parameter 111, the EMA weighting parameter 131, or the EMA coefficient 132 is updated. Suppose.

Now, referring to FIG. 2 again, the processing of the double operation unit 133 will be described.

Referring to FIG. 2, a configuration example of the double operation unit 133 is shown. In the example shown in FIG. 2, among the multiple layers of the verification NN 130 used by the double operation unit 133, the i-th layer (where i=1, 2, . . . , n) is a convolutional layer. , and its convolutional layer is expressed as “f′ _i ”. A weight parameter corresponding to the _i -th layer among the EMA weight parameters 131 is set in the convolution layer f'i. The convolutional layer f′ _i and the convolutional layer f _i may have different set weight parameters, but may perform the same processing.

Also, the input to the convolutional layer _f'i is expressed as _Xi , and the output from the convolutional layer _f'i is expressed as _Z'i . However, the i-th layer of the verification NN 130 may not be a convolutional layer. Here, it is assumed that the input _X1 to the first layer of the verification NN130 is the same as the input _x1 to the first layer of the learning NN110. However, the input x ₁ to the first layer of verification NN 130 may be different than the input x ₁ to the first layer of training NN 110 . That is, the input data to the learning NN 110 and the input data to the verification NN 130 may not be the same.

In addition, the convolution layer f _i included in the learning NN 110 is also used by the double operation unit 133 . Referring to FIG. 2, the input to the convolutional layer f _i used by the double operator 133 is denoted X _i and the output from that convolutional layer f′ _i is denoted Z _i . The dual operation unit 133 calculates the output Z′ _i =f′ _i (X _i ) from the convolutional layer f′ _i , the output Z _i =f _i (X _i ) from the convolutional layer f _i , and the EMA coefficient 132 as Based on a certain α _t , the output Y _i is calculated.

More specifically, the double operation unit 133 performs a weighted summation of the outputs Z _i =f _i (X _i ) and Z′ _i =f′ _i (X _i ) by the EMA coefficient 132 α _t to obtain Calculate the output _Yi . That is, the double operation unit 133 multiplies Z _i =f _i (X _i ) by (1−α _t ), multiplies the output Z′ _i =f′ _i (X _i ) by α _t , and the multiplication result The output Y _i is calculated by adding them together. A formula for calculating the output _Yi is expressed by the following formula (2).

Equation (2) is computed for each corresponding weight parameter in f _i and f′ _i . The input X ₁ to the first layer is the input data contained in the mini-batch, and the input X _i (i≧2) to the second and subsequent layers is the output Y _i−1 from the immediately preceding layer. The double operation unit 133 outputs the output _Yn corresponding to the final layer to the EMA evaluation unit 141 as the EMA inference result. Note that the format of the EMA inference result output from the double computing unit 133 may be the same as the format of the inference result output from the computing unit 112 .

Returning to FIG. 1, the description continues.

(EMA evaluation unit 141)
The EMA evaluation unit 141 obtains evaluation results (hereinafter also referred to as “EMA evaluation results”) based on the EMA inference results for each mini-batch. The EMA evaluator 141 can correspond to an example of a second evaluator. An EMA evaluation result may correspond to an example of a second evaluation result. More specifically, the EMA evaluation unit 141 calculates the loss based on the EMA inference result and the target for each mini-batch, and obtains the EMA evaluation result based on the loss for each mini-batch. The EMA evaluation unit 141 outputs the EMA evaluation result to the EMA updating unit 142 .

Here, the goal is not limited to a specific goal, and a goal similar to that used in general neural networks may be used. For example, when supervised learning is performed, the goal may be the supervised label corresponding to the input data.

The loss function used to calculate the loss is not limited to a specific function, and loss functions similar to loss functions used in general neural networks may be used. For example, the loss function may be the mean squared error based on the difference between the teacher label and the EMA inference result, or the cross entropy error based on the difference between the teacher label and the EMA inference result.

Alternatively, the loss function may be the cross-entropy error based on the difference between the teacher label to which Label Smoothing is applied and the EMA inference result, as shown in Equation (3) below.

Here, l _ema is the loss function based on the EMA inference result, C is the number of classes in the classification problem, δ is the amount of noise, and q is the correct class as 1 and the others as 0. - is the teacher label represented by the hot vector. q _ls is the teacher label with Label Smoothing applied.

(EMA update unit 142)
The EMA updating unit 142 updates the EMA coefficients 132 based on the EMA evaluation results output from the EMA evaluating unit 141 . This allows the EMA weight parameter 131 to be trained such that the EMA inference result approaches the teacher label. The EMA updater 142 may correspond to an example of a weighting factor updater. More specifically, the EMA updating unit 142 may update the weighting factor _αt , which is the EMA coefficient 132, by error backpropagation based on the EMA evaluation result. As an example, the weighting factor α _t update formula can be expressed by the following formula (4).

where l _ema is the loss function based on the EMA inference results and μ is the learning rate. A condition for updating the weighting factor _αt , which is the EMA coefficient 132, may be set, and the weighting factor _αt may be updated when the update condition is satisfied. For example, the update condition may be a condition that the weight parameter 111 has been updated, or a condition that an update timing specified in advance by the user has arrived.

As the update timing, an update interval (for example, once every predetermined number of times) may be specified. Alternatively, the update interval of the weighting factor _αt may be always equal intervals from the start of updating the weighting parameter 111, but may not be equal intervals. For example, the weighting factor α _t may not be updated in the initial stage immediately after the weighting parameter 111 is updated (for example, until the weighting parameter 111 is updated a predetermined number of times).

(storage unit 143)
The storage unit 143 updates the EMA weighting parameter 131 based on the updated weighting parameter 111 , the EMA weighting parameter 131 , and the updated EMA coefficient 132 . The storage unit 143 can correspond to an example of a second updating unit.

More specifically, the storage unit 143 updates the EMA weight parameter 131 by the weighted sum of the updated weight parameter 111 and the EMA weight parameter 131 by the updated EMA coefficient 132 . That is, the storage unit 143 multiplies the updated weight parameter 111 by (1−α _t+1 ), multiplies the EMA weight parameter 131 by α _t+1 , and adds the multiplication results together to obtain the updated EMA weight parameter 131 is calculated. A formula for calculating the EMA weighting parameter 131 after such updating is expressed by the following formula (5).

As described above, learning by the learning device 10 updates the weighting parameter 111 and the EMA weighting parameter 131 .

The configuration example of the learning device according to the first embodiment of the present invention has been described above.

(Operation in learning stage)
Next, with reference to FIG. 3, the flow of operations in the "learning stage" performed by the learning device 10 according to the first embodiment of the present invention will be described. FIG. 3 is a flow chart showing an operation example of the learning stage performed by the learning device 10 according to the first embodiment of the present invention.

First, the input unit 102 creates a mini-batch by obtaining batch size input data from the learning data set 101, and outputs the created mini-batch to the calculation unit 112 of the learning NN 110 (S101).

Subsequently, the calculation unit 112 obtains an inference result corresponding to the mini-batch based on the mini-batch created by the input unit 102 and the learning NN 110 to which the weight parameter 111 is set (S102). The calculation unit 112 outputs the inference result corresponding to the mini-batch to the evaluation unit 121 .

The evaluation unit 121 obtains an evaluation result by evaluating the inference result output from the calculation unit 112 . The evaluation unit 121 outputs the evaluation result to the update unit 122 (S103). The update unit 122 updates the weight parameter 111 of the learning NN 110 based on the evaluation result output from the evaluation unit 121 (S104). If verification NN 110 is to be updated ("NO" in S105), that is, if the update condition for EMA coefficient 132 (weighting coefficient α _t ) is not satisfied, the operation proceeds to S110.

On the other hand, when updating the verification NN 110 (“YES” in S105), that is, when the update condition of the EMA coefficient 132 (weighting coefficient α _t ) is satisfied, the double operation unit 133 receives the mini-batch from the input unit 102. input. Then, the double computing unit 133 receives the mini-batch input from the input unit 102, the learning NN 110 to which the updated weighting parameter 111 is set, the verification NN 130 to which the EMA weighting parameter 131 is set, and the EMA coefficient 132. , the EMA inference result is obtained using equation (2). The double operation unit 133 outputs the EMA inference result to the EMA evaluation unit 141 (S106).

The EMA evaluation unit 141 obtains an EMA evaluation result by evaluating the EMA inference result output from the double operation unit 133 (S107). The EMA evaluation unit 141 outputs the EMA evaluation result to the EMA updating unit 142 . The EMA updating unit 142 updates the weighting factor _αt , which is the EMA coefficient 132, based on the EMA evaluation result output from the EMA evaluating unit 141 (S108).

Thereafter, the storage unit 143 updates the EMA weighting parameter 131 using Equation (5) based on the updated weighting parameter 111, the EMA weighting parameter 131, and the updated EMA coefficient 132 (S109). .

If the learning end condition of the learning NN 110 is not satisfied ("NO" in S110), the operation proceeds to S101. On the other hand, if the learning termination condition of learning NN 110 is satisfied ("YES" in S110), learning is terminated.

The flow of operations in the "learning stage" performed by the learning device 10 according to the first embodiment of the present invention has been described above.

(Summary of the first embodiment)
As described above, the learning apparatus 10 according to the first embodiment of the present invention performs the weighted sum of the weight parameters of the verification NN and the weight parameters of the learning NN in the EMA processing in constructing the verification NN. Make it possible to learn the coefficients used when calculating As a result, the optimum coefficient is automatically determined, so that the effect of enabling stable learning can be enjoyed. In addition, this provides the effect that the user's burden of adjusting the coefficients can be reduced.

The first embodiment of the present invention has been described above.

(2. Second embodiment)
Next, a second embodiment of the invention will be described. In the first embodiment of the present invention, the case where the EMA coefficient 132 is the scalar coefficient α _t has been described. This means that all layers building the NN are weighted by the same value. In the second embodiment of the present invention, a case will be described in which an EMA coefficient is provided for each layer included in the NN.

A functional configuration diagram of the learning device 10 according to the second embodiment of the present invention is the same as the functional configuration diagram (FIG. 1) of the learning device 10 according to the first embodiment of the present invention. Further, the learning device 10 according to the second embodiment of the present invention has an EMA coefficient 132, a double operation unit 133, an EMA updating unit 142 and The functions of the storage unit 143 are different, and the functions of the other components are the same.

Therefore, hereinafter, the EMA coefficient 132, the double operation unit 133, the EMA update unit 142, and the storage unit 143, which are included in the learning device 10 according to the second embodiment of the present invention, will be mainly described, and other components will be described. A detailed description of the functions it has is omitted.

In the second embodiment of the present invention, the EMA coefficients 132 are represented by weighting coefficients α _t,i (i=1, 2, . . . , n). That is, in the second embodiment of the present invention, the number of weighting coefficients α _t that are the EMA coefficients 132 is the same as the total number n of layers that constitute the learning NN 110 or the verification NN 130 .

(Double operation unit 133)
The double computing unit 133 includes the input data included in the mini-batch output from the input unit 102, the learning NN 110 set with the weighting parameter 111 updated by the updating unit 122, and the verification NN 130 provided for each layer. EMA inference results are obtained based on the calculated EMA coefficients 132 .

FIG. 4 is a diagram for explaining the processing of the double computing unit 133 according to the second embodiment of the present invention. Referring to FIG. 4, a configuration example of the double operation unit 133 is shown. In the example shown in FIG. 4, attention is paid to the kth layer and the (k+1)th layer (where k=1, 2, . . . , n−1). Note that the convolution layer f _k has the same configuration as the convolution layer f _i according to the first embodiment of the present invention, and the convolution layer f′ _k has the same configuration as the convolution layer f k according to the first embodiment of the present invention. ' has the same configuration as _i .

The input to the convolutional layer _f'k is expressed as _Xk , and the output from the convolutional layer _f'k is expressed as _Z'k . Also, the input to the convolutional layer _fk is expressed as _Xk , and the output from the convolutional layer _f'k is expressed as _Zk . The dual operation unit 133 outputs Z′ _k =f′ _k (X _k ) from the convolutional layer f′ _k , Z _k =f _k (X _k ) from the convolutional layer f _k , and the k-th layer Based on the corresponding EMA coefficients 132, α _t,k , the output Y _k is calculated.

_{More specifically ,} the _dual operation unit 133 _{calculates α t , k} to compute the output Y _k . That is, the double operation unit 133 multiplies Z _k =f _k (X _k ) by (1−α _t,k ), and multiplies the output Z′ _k =f′ _k (X _k ) by α _t,k Then, the output _Yk is calculated by adding the multiplication results together. A formula for calculating the output _Yk is expressed by the following formula (6).

Note that the output Y _k becomes the input X _k+1 to the (k+1)th layer. Processing in the (k+1)-th layer is also performed in the same manner as in the k-th layer. That is, the dual operation unit 133 outputs Z′ _k+1 =f′ _k+1 (X _{k+1 )} from the convolutional layer f′ _k+1 , Z _k+1 =f _k+1 (X _k+1 ) from the convolutional layer f k+1 , and the EMA coefficient Based on α _t,k+1 which is 132, the output Y _k+1 is calculated.

More specifically, the double operation unit 133 performs a weighted summation of the output Z′ _k+1 =f′ _k+1 (X _k+1 ) and the output Z _k+1 =f _k+1 (X _k+1 ) by α _t,k+1 which is the EMA coefficient 132. to calculate the output Yk ₊₁ . That is, the double operation unit 133 multiplies Z _k+1 =f _k+1 (X _k+1 ) by (1−α _t,k+1 ), and multiplies the output Z′ _k+1 =f′ _k+1 (X _k+1 ) by α _t,k+1 Then, the output Yk ₊₁ is calculated by adding the multiplication results together.

Note that the output Y _k+1 becomes the input X _k+2 to the (k+2)th layer. In this way, each layer from the first layer to the final layer is processed, and the output _Yn corresponding to the final layer is calculated. The double operation unit 133 outputs the output _Yn corresponding to the final layer to the EMA evaluation unit 141 as the EMA inference result.

(EMA update unit 142)
The EMA updating unit 142 updates the weighting coefficients α _t,k (k=1, 2, . This allows the EMA weight parameter 131 to be trained such that the EMA inference result approaches the teacher label. More specifically, the EMA updating unit 142 updates weighting coefficients α _t,k (k=n, n−1, . . . , 1 ) may be updated sequentially. As an example, the weighting coefficient α _t,k update formula can be expressed by the following formula (7).

(storage unit 143)
For each layer from the first layer to the last layer, the storage unit 143 multiplies the updated weight parameter 111 by (1−α _t+1,k ), multiplies the EMA weight parameter 131 by α _t+1,k , and stores the multiplication result The updated EMA weight parameter 131 is calculated by adding them together. A formula for calculating the EMA weighting parameter 131 after such updating is expressed by the following formula (8).

As described above, learning by the learning device 10 updates the weighting parameter 111 and the EMA weighting parameter 131 .

The configuration example of the learning device according to the second embodiment of the present invention has been described above.

(Operation in learning stage)
Next, the flow of operations in the "learning stage" performed by the learning device 10 according to the second embodiment of the present invention will be described. A flowchart showing an operation example of the "learning phase" executed by the learning device 10 according to the second embodiment of the present invention is a "learning phase" executed by the learning device 10 according to the first embodiment of the present invention. This is the same as the flow chart (FIG. 2) showing an example of the operation of .

In addition, the learning device 10 according to the second embodiment of the present invention has double operation unit 133, EMA updating unit 142, and storage unit 143, compared to learning device 10 according to the first embodiment of the present invention. Operational details differ, and other operational details are similar. Therefore, hereinafter, the operations of the double operation unit 133, the EMA update unit 142, and the storage unit 143 of the learning device 10 according to the second embodiment of the present invention will be mainly described, and other operations will be described in detail. Description is omitted.

Also in the second embodiment of the present invention, S101 to S105 are executed as in the first embodiment of the present invention.

Next, the double computing unit 133 receives the mini-batch input from the input unit 102, the learning NN 110 set with the updated weighting parameter 111, the verification NN 130 with the EMA weighting parameter 131 set, and the first layer (6) is used to obtain the EMA inference result based on the EMA coefficients 132 corresponding to each layer from to the final layer. The double operation unit 133 outputs the EMA inference result to the EMA evaluation unit 141 (S106).

The EMA evaluation unit 141 obtains an EMA evaluation result by evaluating the EMA inference result output from the double operation unit 133 (S107). The EMA evaluation unit 141 outputs the EMA evaluation result to the EMA updating unit 142 . The EMA update unit 142 updates the EMA coefficients 132 corresponding to each layer from the first layer to the final layer using Equation (7) based on the EMA evaluation results output from the EMA evaluation unit 141 (S108).

After that, the storage unit 143 uses Equation (8) based on the updated weight parameter 111, the EMA weight parameter 131, and the updated EMA coefficient 132 corresponding to each layer from the first layer to the final layer. Then, the EMA weight parameter 131 is updated (S109).

Also in the second embodiment of the present invention, S110 is executed as in the first embodiment of the present invention.

The flow of operations in the "learning stage" performed by the learning device 10 according to the second embodiment of the present invention has been described above.

(Summary of the second embodiment)
As described above, according to the second embodiment of the present invention, the weighted sum of the weight parameter of the verification NN and the weight parameter of the learning NN is calculated in the EMA processing in constructing the verification NN. Not only are the sometimes used coefficients learnable, but the coefficients are provided for each layer of the training NN. As a result, the optimum coefficient is automatically determined for each layer, so not only can the same effects as in the first embodiment of the present invention be obtained, but a verification NN with a higher degree of freedom can be constructed. .

The second embodiment of the present invention has been described above.

(3. Hardware configuration example)
Next, a hardware configuration example of the learning device 10 according to the first embodiment of the present invention will be described. Note that the hardware configuration of the learning device 10 according to the second embodiment of the present invention can also be realized in the same manner as the hardware configuration of the learning device 10 according to the first embodiment of the present invention.

A hardware configuration example of the information processing device 900 will be described below as a hardware configuration example of the learning device 10 according to the first embodiment of the present invention. Note that the hardware configuration example of the information processing device 900 described below is merely an example of the hardware configuration of the learning device 10 . Therefore, as for the hardware configuration of the learning device 10, unnecessary configurations may be deleted from the hardware configuration of the information processing device 900 described below, or a new configuration may be added.

FIG. 5 is a diagram showing the hardware configuration of an information processing device 900 as an example of the learning device 10 according to the first embodiment of the present invention. The information processing device 900 includes a CPU (Central Processing Unit) 901, a ROM (Read Only Memory) 902, a RAM (Random Access Memory) 903, a host bus 904, a bridge 905, an external bus 906, and an interface 907. , an input device 908 , an output device 909 , a storage device 910 and a communication device 911 .

The CPU 901 functions as an arithmetic processing device and a control device, and controls general operations within the information processing device 900 according to various programs. Alternatively, the CPU 901 may be a microprocessor. The ROM 902 stores programs, calculation parameters, and the like used by the CPU 901 . The RAM 903 temporarily stores programs used in the execution of the CPU 901, parameters that change as appropriate during the execution, and the like. These are interconnected by a host bus 904 comprising a CPU bus or the like.

The host bus 904 is connected via a bridge 905 to an external bus 906 such as a PCI (Peripheral Component Interconnect/Interface) bus. Note that the host bus 904, the bridge 905 and the external bus 906 do not necessarily have to be configured separately, and these functions may be implemented in one bus.

The input device 908 includes input means for the user to input information, such as a mouse, keyboard, touch panel, button, microphone, switch, and lever, and an input control circuit that generates an input signal based on the user's input and outputs it to the CPU 901 . etc. A user who operates the information processing apparatus 900 can input various data to the information processing apparatus 900 and instruct processing operations by operating the input device 908 .

The output device 909 includes, for example, a CRT (Cathode Ray Tube) display device, a liquid crystal display (LCD) device, an OLED (Organic Light Emitting Diode) device, a display device such as a lamp, and an audio output device such as a speaker.

The storage device 910 is a device for data storage. The storage device 910 may include a storage medium, a recording device that records data on the storage medium, a reading device that reads data from the storage medium, a deletion device that deletes data recorded on the storage medium, and the like. The storage device 910 is configured by, for example, an HDD (Hard Disk Drive). The storage device 910 drives a hard disk and stores programs executed by the CPU 901 and various data.

The communication device 911 is, for example, a communication interface configured with a communication device or the like for connecting to a network. Also, the communication device 911 may support either wireless communication or wired communication.

The hardware configuration example of the learning device 10 according to the first embodiment of the present invention has been described above.

(4. Summary)
Although the preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, the present invention is not limited to such examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention belongs can conceive of various modifications or modifications within the scope of the technical idea described in the claims. It is understood that these also naturally belong to the technical scope of the present invention.

In the first embodiment of the present invention and the second embodiment of the present invention, the case where the learning data is image data (particularly, the case where it is still image data) has been mainly described. However, the type of learning data is not particularly limited. For example, data other than still image data can be used as learning data if a feature amount suitable for the type of learning data is extracted. For example, the learning data may be moving image data including a plurality of frames, audio data, or other time-series data.

At this time, when the learning data is still image data, generally two-dimensional convolution layers are used as the convolution layers included in each of the learning NN 110 and the verification NN 130 . On the other hand, if a three-dimensional convolutional layer is used as the convolutional layer included in each of the learning NN 110 and the verification NN 130, moving image data can be applied as the learning data.

In the first embodiment of the invention and the second embodiment of the invention, no restriction is placed on the EMA factor 132, ie the weighting factor α. However, a function that maps to the range (0, 1) may be applied to the weighting factor α. For example, the function mapped to the range (0,1) may be the sigmoid function sgm. As a result, it can be expected that more stable learning of the weighting factor α will be performed. If a sigmoid function is applied to α, equations (2) and (5) are replaced by equations (9) and (10) below.

10 learning device 101 learning data set 102 input unit 111 weight parameter 112 calculation unit 121 evaluation unit 122 update unit 131 EMA weight parameter 132 EMA coefficient 133 double calculation unit 141 EMA evaluation unit 142 EMA update unit 143 storage unit

Claims (11)

a first computing unit that outputs a first inference result of the first input data based on the first input data and a first neural network;
a first evaluation unit that outputs a first evaluation result based on the first inference result;
a first updating unit that updates a first weight parameter of the first neural network based on the first evaluation result;
Based on the second input data, the first neural network set with the updated first weight parameter, the second neural network, and the weighting factor of the second neural network, the a second computing unit that outputs a second inference result of the second input data;
a second evaluation unit that outputs a second evaluation result based on the second inference result;
a weighting factor updating unit that updates the weighting factor based on the second evaluation result;
A second updating unit that updates the second weighting parameter based on the updated first weighting parameter, the second weighting parameter of the second neural network, and the updated weighting factor. and,
An information processing device. The second computing unit
inputting the second input data to the first layer of each of the first neural network and the second neural network in which the updated first weight parameter is set;
The information processing device according to claim 1 . The second computing unit
For each of the first neural network and the second neural network to which the updated first weight parameter is set, the second and subsequent layers are weighted by the weight coefficient with respect to the output from the immediately preceding layer. enter the sum,
The information processing apparatus according to claim 1 or 2. The second evaluation unit calculates a loss based on the second evaluation result and the teacher label, and outputs the second evaluation result based on the loss.
The information processing apparatus according to any one of claims 1 to 3. The weighting factor updating unit
updating the weighting factor by error backpropagation based on the second evaluation result;
The information processing apparatus according to any one of claims 1 to 4. The second updating unit
updating the second weighting parameter by a weighted sum of the updated first weighting parameter and the second weighting parameter of the second neural network by the updated weighting factor;
The information processing apparatus according to any one of claims 1 to 5. The weighting factor is provided for each layer included in the first neural network and the second neural network.
The information processing apparatus according to any one of claims 1 to 6. The first input data and the second input data are the same data,
The information processing apparatus according to any one of claims 1 to 7. The first input data and the second input data are different data,
The information processing apparatus according to any one of claims 1 to 7. outputting a first inference result of the first input data based on the first input data and a first neural network;
outputting a first evaluation result based on the first inference result;
updating a first weight parameter of the first neural network based on the first evaluation result;
Based on the second input data, the first neural network set with the updated first weight parameter, the second neural network, and the weighting factor of the second neural network, the outputting a second inference result for the second input data;
outputting a second evaluation result based on the second inference result;
updating the weighting factor based on the second evaluation result;
updating the second weighting parameter based on the updated first weighting parameter, the second weighting parameter of the second neural network, and the updated weighting factor;
A method of processing information, comprising: the computer,
a first computing unit that outputs a first inference result of the first input data based on the first input data and a first neural network;
a first evaluation unit that outputs a first evaluation result based on the first inference result;
a first updating unit that updates a first weight parameter of the first neural network based on the first evaluation result;
Based on the second input data, the first neural network set with the updated first weight parameter, the second neural network, and the weighting factor of the second neural network, the a second computing unit that outputs a second inference result of the second input data;
a second evaluation unit that outputs a second evaluation result based on the second inference result;
a weighting factor updating unit that updates the weighting factor based on the second evaluation result;
A second updating unit that updates the second weighting parameter based on the updated first weighting parameter, the second weighting parameter of the second neural network, and the updated weighting factor. and,
A program that functions as an information processing device comprising

Images