JP7059695B2 - Learning method and learning device - Google Patents

Description

The present invention relates to a learning method and a learning device for machine learning a recognition processing unit that extracts features from input data and performs identification using the extracted features.

Conventionally, machine learning of a recognition processing unit that performs recognition processing on recognition target data has been generally performed. As one method of such machine learning, for example, after learning the recognition processing unit using data with a correct answer prepared in advance (pre-learning data), at a site (for example, a store) where the recognition processing unit is used. There is a method of re-learning (additional learning) of the recognition processing unit using the obtained data (local data). This method is a method of improving the adaptability of the recognition processing unit to not only the pre-learning data but also the local data by performing additional learning to improve the generalization. Note that generalization is the ability of a machine-learned computer (for example, a recognition processing unit) to derive a correct answer (output) for an unlearned question (input) similar to a learned question, that is. For input data similar to the trained data, it refers to the ability to output results similar to the results of the trained data. An example of performing additional learning using local data after performing pre-learning as described above is also disclosed in, for example, Patent Document 1.

Japanese Unexamined Patent Publication No. 2016-143353 (see claim 1, paragraph [0005], [0006], [0018], [0021] to [0043], FIG. 1 and the like).

However, in the conventional method of performing pre-learning and additional learning of the recognition processing unit, the performance of the recognition processing unit after the additional learning is not verified at all, so whether or not the performance of the recognition processing unit is really improved by the additional learning. It is unknown. For example, when the additional learning becomes overfitting, the recognition processing unit is adapted only to the actually learned data and is not adapted to the unknown data, so that the generalization property is lowered. If the performance of the recognition processing unit is not verified, it is not possible to detect the deterioration in the performance of the recognition processing unit due to this overfitting. Therefore, if the parameters (for example, weights) of the recognition processing unit are unconditionally changed (updated) by additional learning, the generalization property may be reduced due to the performance deterioration of the recognition processing unit.

The present invention has been made to solve the above problems, and an object thereof is to verify the performance of the recognition processing unit by additional learning and then update the parameters even when pre-learning and additional learning are performed. It is an object of the present invention to provide a learning method and a learning device capable of improving generalization by avoiding deterioration of the performance of the recognition processing unit due to additional learning.

The learning method according to one aspect of the present invention is a learning method in machine learning of a recognition processing unit that extracts features from input data and identifies using the extracted features, and is data with a correct answer obtained in advance. There is no correct answer acquired in the pre-learning step (1) in which the pre-learning of the recognition processing unit is performed using the set A, and in an environment different from the data set A and the data set A after the pre-learning. The additional learning step (2) for performing additional learning of the recognition processing unit, the performance evaluation step (3) for evaluating the performance of the recognition processing unit by the additional learning, and the performance evaluation using the data set B of the above. Based on the result of the above, the update step (4) of updating the parameters of the recognition processing unit to the parameters obtained by the additional learning is included.

The learning device according to another aspect of the present invention includes a recognition processing unit that extracts features from input data and performs identification using the extracted features, and a learning control unit that causes the recognition processing unit to perform machine learning. The learning control unit was acquired in an environment different from the data set A and the data set A for the recognition processing unit that had been pre-learned using the data set A with the correct answer obtained in advance. After additional learning using the data set B without a correct answer, the performance evaluation of the recognition processing unit is performed by the additional learning, and the parameters of the recognition processing unit are subjected to the additional learning based on the result of the performance evaluation. Update to the parameters obtained by.

According to the above learning method and learning device, not only pre-learning and additional learning of the recognition processing unit are performed, but also performance evaluation by additional learning is performed, and parameters are updated based on the results. As a result, it is possible to avoid deterioration in the performance of the recognition processing unit (for example, due to overfitting) and improve generalization.

It is a block diagram which shows the schematic structure of the learning apparatus of embodiment of this invention. It is explanatory drawing which shows typically one configuration example of the recognition processing part of the said learning apparatus. It is explanatory drawing which shows typically the unit which constitutes each layer of the recognition processing part. It is a flowchart which shows the process flow of the learning method using the said learning apparatus. It is explanatory drawing which shows typically the data propagation model at the time of the pre-learning of the recognition processing part. It is explanatory drawing which shows typically the data propagation model at the time of the additional learning of the feature extraction layer of the recognition processing part. It is explanatory drawing which shows typically the data propagation model at the time of the performance evaluation of the said feature extraction layer. It is explanatory drawing which shows typically the data propagation model at the time of the additional learning of the identification layer of the recognition processing part. It is explanatory drawing which shows typically the data propagation model at the time of performance evaluation of the said identification layer. It is explanatory drawing which shows the data propagation model at the time of actually performing recognition in the learning apparatus after learning by the said learning method. It is a flowchart which shows the flow of the other processing of the said learning method.

An embodiment of the present invention will be described below with reference to the drawings. The present invention is not limited to the following contents.

[About learning device]
FIG. 1 is a block diagram showing a schematic configuration of the learning device 1 of the present embodiment. The learning device 1 is a device capable of machine learning a processing unit (for example, corresponding to a recognition processing unit 11 described later) that performs predetermined processing on input data, and is a terminal such as a personal computer (PC). It consists of devices. Specifically, the learning device 1 has a recognition processing unit 11, a storage unit 12, an input unit 13, a display unit 14, a communication unit 15, and a control unit 16.

The control unit 16 is composed of a central processing unit (CPU) that controls the operation of each unit of the learning device 1, and follows an operation program stored in the storage unit 12 (program storage unit 12a described later). Operate. In the present embodiment, the control unit 16 functions as a learning control unit that causes the recognition processing unit 11 to perform machine learning.

The recognition processing unit 11 is an arithmetic unit that extracts features from input data and performs recognition using the extracted features. For example, the recognition processing unit 11 is composed of a GPU (Graphics Processing Unit) that is an arithmetic unit specialized in real-time image processing. ing. The recognition processing unit 11 may be configured by the same CPU as the control unit 16 or by a separate CPU.

Here, as the input data, for example, image data (video or still image data) acquired by shooting with an image pickup unit 2 (camera) installed in a store can be considered. The above image data is input to the learning device 1 in real time from the image pickup unit 2, for example, via a wired or wireless communication line (for example, a LAN (Local Area Network) or the Internet). The above image data is temporarily recorded on a portable recording medium (optical disk, non-volatile memory, etc.), and the image data recorded on the recording medium is read by a reading device (not shown) to read the input data. May be taken into the learning device 1. The details of the recognition processing unit 11 will be described later.

The storage unit 12 has a program storage unit 12a, a learning / evaluation data storage unit 12b, and a processing data storage unit 12c. The program storage unit 12a is a memory for storing an operation program for operating each unit of the learning device 1. The learning / evaluation data storage unit 12b stores learning data (for example, data set A described later) and data for performance (quality) evaluation (for example, data set C described later) used in machine learning of the recognition processing unit 11. It is a memory to store. The processing data storage unit 12c is acquired in the field where the image data (for example, data set B described later) input to the learning device 1 from the outside (for example, the imaging unit 2), that is, the recognition processing of the recognition processing unit 11 is utilized. This is a memory for temporarily storing the data input to the learning device 1. The storage unit 12 (program storage unit 12a, learning / evaluation data storage unit 12b, processing data storage unit 12c) is composed of, for example, a hard disk, but in addition, RAM (Random Access Memory) and ROM (Read Only Memory). , An optical disk, a photomagnetic disk, a recording medium such as a non-volatile memory, and the like may be appropriately selected and configured.

The input unit 13 is composed of, for example, a keyboard, a mouse, a touch pad, a touch panel, or the like, and receives various instruction inputs by an operator who operates the learning device 1, for example. The display unit 14 is a device that displays various information including the recognition result of the recognition processing unit 11, and is composed of, for example, a liquid crystal display device. The communication unit 15 is an interface for communicating with an external terminal (for example, a mobile terminal such as a smartphone or another PC), and includes an antenna, a transmission / reception circuit, a modulation circuit, a demodulation circuit, and the like.

[Details of the recognition processing unit]
Next, the details of the recognition processing unit 11 described above will be described. As shown in FIG. 2, the recognition processing unit 11 is composed of a hierarchical neural network having an input layer 11a, a feature extraction layer 11b, and an identification layer 11c. A neural network is an information processing system that imitates a human neural network. In a neural network, an engineering neuron model corresponding to a nerve cell is referred to here as a unit U. The input layer 11a, the feature extraction layer 11b, and the identification layer 11c each have a plurality of units U.

The input layer 11a usually consists of one layer. For example, image data (pixel values) of each pixel constituting one image is input to each unit U of the input layer 11a. The image data is output as it is from each unit U of the input layer 11a to the feature extraction layer 11b.

The feature extraction layer 11b is a layer that extracts features from the input data (data of the input layer 11a) and outputs the features to the identification layer 11c. In the feature extraction layer 11b, for example, a process of extracting a region in which a person is reflected from an input image is performed by an operation in each unit U.

The identification layer 11c is a layer for performing identification using the features extracted by the feature extraction layer 11b. In the identification layer 11c, for example, a process of identifying the posture or behavior of the person is performed from the area of the person extracted by the feature extraction layer 11b by the calculation in each unit U.

The feature extraction layer 11b and the identification layer 11c are each composed of a plurality of layers. As a result, the deep neural network (DNN) is configured as the entire recognition processing unit 11.

Here, the description of each unit U constituting the feature extraction layer 11b and the identification layer 11c will be supplemented. As the unit U, an element having multiple inputs and one output, as shown in FIG. 3, is usually used. The signal is transmitted in only one direction, is given a certain weight (coupling load: wi), and is input to the unit U. This weight represents the strength of the bond between the units U and U arranged hierarchically, and the above weight can be changed by learning. The value X obtained by subtracting the threshold value θ from the sum of each weighted input value (wixi) is transformed by the response function f (X) and then output from the unit U. That is, the output value y of the unit U is expressed by the following mathematical formula.
y = f (X)
here,
X = Σ (wixi-θ)
Is. As the response function, for example, a sigmoid function can be used.

Each unit U of the input layer 11a usually does not have a sigmoid characteristic or a threshold value, and therefore the input value appears as it is in the output as described above. On the other hand, each unit U of the final layer (output layer) of the identification layer 11c outputs the identification result (data corresponding to the identified posture and behavior) in the identification layer 11c.

As a learning algorithm of the discrimination processing unit 11 (hierarchical neural network), for example, the steepest descent so that the square error between the value indicating the correct answer (data) and the output value (data) from the discrimination layer 11c is minimized. An error backpropagation method (backpropagation) is used in which the weight of the identification layer 11c and the weight of the feature extraction layer 11b are sequentially changed by a method.

[About learning method]
Next, the learning method in the learning device 1 having the above configuration will be described. FIG. 4 is a flowchart showing a processing flow of the learning method of the present embodiment. The learning method of the present embodiment is a learning method in machine learning of the recognition processing unit 11, and can be broadly viewed as a pre-learning step (I), a first learning / evaluation / updating step (II), and a second. Including the learning / evaluation / updating process (III) of. Each of these steps is sequentially performed under the control of the control unit 16. Hereinafter, details of each step will be described.

<Pre-learning process>
The pre-learning step (I) is a step (S1) of performing pre-learning (offline learning) of the recognition processing unit 11. In the step of S1, the recognition processing unit 11 is pre-learned using the data set A with the correct answer (with supervised and annotation) obtained in advance. Here, the data set corresponds to a plurality of sets (a plurality of images) of the image data, with the image data (pixel values) of each pixel constituting one image as one set. The image data included in the above data set A includes, for example, a place (for example, a studio away from the store) different from the store (site) when the recognition processing unit 11 performs customer behavior recognition processing in the store. Etc.), and the image data acquired in advance can be considered. The data set A is stored in advance in the learning / evaluation data storage unit 12b, and pre-learning is performed using this data set A.

FIG. 5 schematically shows a data propagation model during pre-learning. The data included in the data set A is input to the input layer 11a of the recognition processing unit 11 for each image, and is output from the input layer 11a to the feature extraction layer 11b. In each unit U of the feature extraction layer 11b, a weighted operation is performed on the input data. By this calculation, the feature extraction layer 11b extracts features (for example, a person's area) from the input data, and the data indicating the extracted features is output to the identification layer 11c.

In each unit U of the identification layer 11c, a weighted operation is performed on the input data, whereby identification based on the above characteristics (for example, analysis of the behavior of a person) is performed. The data indicating the identification result is output from the identification layer 11c.

The control unit 16 compares the output value (data) of the identification layer 11c with the value indicating the correct answer, calculates these errors (losses), and extracts the weights and features of the identification layer 11c so that this error becomes small. The weights of the layers 11b are sequentially changed (backpropagation). As a result, the identification layer 11c and the feature extraction layer 11b are learned.

<First learning / evaluation / renewal price process>
The first learning / evaluation / updating step (II) includes a step (S2) of performing additional learning L1 of the feature extraction layer 11b, a step (S3) of performing performance evaluation E1 of the feature extraction layer 11b by the additional learning L1. A step (S4 to S6) of updating the parameters (weights) of the feature extraction layer 11b to the parameters obtained by the additional learning L1 based on the result of the performance evaluation E1 is included.

(Additional learning L1)
In S2, after the pre-learning step (I), the additional learning L1 of the feature extraction layer 11b is performed by using the data set A with the correct answer and the data set B without the correct answer acquired in an environment different from the data set A. I do. Here, as the above-mentioned data set B, a data set to be recognized by the recognition processing unit 11 can be considered. For example, when the recognition processing unit 11 performs customer behavior recognition processing in the store. Can consider a set of image data (image data acquired by the image pickup unit 2 in the store) acquired at the store (site) and input to the learning device 1. The data set B is stored in the processing data storage unit 12c, and is used for the additional learning L1 together with the above-mentioned data set A previously stored in the learning / evaluation data storage unit 12b. For convenience of explanation below, among the data sets B, the data set used for the additional learning L1 is referred to as a data set B1.

FIG. 6 schematically shows a data propagation model during the additional learning L1 of the feature extraction layer 11b. In S2, the identification layer 11c'is used instead of the identification layer 11c. The identification layer 11c'is an identification layer used only when the additional learning L1 of the feature extraction layer 11b, the performance evaluation E1 described later, and the parameter update are performed, and is composed of a temporary identification layer or a decoder. The tentative discrimination layer is a discrimination layer configured for learning with weak supervised learning. On the other hand, the decoder is a layer configured for unsupervised learning, and performs a process (decoding) of returning the output value from the feature extraction layer 11b to an input value. That is, when learning with weak supervised learning is performed, a temporary identification layer is used as the discrimination layer 11c', and when learning without supervised learning is performed, a decoder is used as the discrimination layer 11c'.

Here, weak supervised learning means that learning is performed by regarding the discrimination result as a correct answer only when the correct answer rate (score) obtained based on the output value output by the discrimination layer 11c'is equal to or higher than a certain threshold value. .. The correct answer rate is an index expressed by a / n when the number of test data is n and the number of correct answers is a.

In S2, the additional learning L1 is performed while the data set A and the data set B are alternately input to the input layer 11a, for example. By performing additional learning L1 using the data set B1 together with the data set A, the feature extraction layer 11b is adapted not only to the data set A but also to the data set B1, that is, to the data obtained in the field. Is possible. When the data set A is used as the input data, the additional learning L1 of the feature extraction layer 11b is performed by the same procedure as the pre-learning of S1.

On the other hand, when the data set B1 is used as the input data, the additional learning L1 of the feature extraction layer 11b is performed as follows. When supervised learning is performed as additional learning L1, the control unit 16 satisfies a certain criterion when the correct answer rate obtained based on the output value of the provisional identification layer as the identification layer 11c'is equal to or more than the threshold value. ), The parameter of the feature extraction layer 11b is not updated, and the error (loss) between the correct answer rate and the threshold value is found only when the correct answer rate is less than the threshold value (only when a certain standard is not satisfied). The parameters of the feature extraction layer 11b are updated (weights are changed) so as to be smaller. Further, when unsupervised learning is performed as the additional learning L1, the control unit 16 determines whether or not the output value of the decoder matches the original data (input value of the input layer 11a), and if they match (constant). (If the criteria are satisfied), the parameters of the feature extraction layer 11b are not updated, and only if they do not match (only if they do not meet certain criteria), the error (loss) between the output value and the input value. The parameters of the feature extraction layer 11b are updated (weights are changed) so that As a result, the feature extraction layer 11b is additionally learned.

(Performance evaluation E1)
In the step of S3, the control unit 16 performs the performance evaluation E1 of the feature extraction layer 11b by the additional learning L1 of S2. More specifically, the control unit 16 performs the performance evaluation E1 of the feature extraction layer 11b by using the data set B2 other than the data set B1 included in the data set B. When performing the performance evaluation E1 of the feature extraction layer 11b, the same identification layer 11c'(temporary identification layer or decoder) used in the additional learning L1 of S2 is used as the identification layer.

FIG. 7 schematically shows a data propagation model at the time of performance evaluation E1 of the feature extraction layer 11b. In S3, the evaluation data set B2, that is, the data not used in the additional learning L1 among the field data is used, and the data of the data set B2 is input to the input layer 11a. The above data input to the input layer 11a is output to the feature extraction layer 11b. In the feature extraction layer 11b, a weighted operation is performed on the input data, and the data indicating the operation result is output to the identification layer 11c'. In the identification layer 11c', a weighted operation is performed on the data input from the feature extraction layer 11b, and the data indicating the operation result is output from the identification layer 11c'.

When learning with weak supervised learning is performed in S2, the control unit 16 determines whether or not the correct answer rate obtained based on the output value of the provisional identification layer as the identification layer 11c'is equal to or higher than a certain threshold value. Performance evaluation E1 of the feature extraction layer 11b is performed. For example, when the correct answer rate is equal to or higher than the threshold value, the control unit 16 determines that the performance of the feature extraction layer 11b satisfies a certain standard. On the other hand, when the correct answer rate is less than the threshold value, the control unit 16 determines that the performance of the feature extraction layer 11b does not satisfy a certain standard.

Further, when unsupervised learning is performed in S2, the control unit 16 determines whether or not the output value of the decoder as the identification layer 11c'matches the original data (input value of the input layer 11a). , Performance evaluation E1 of the feature extraction layer 11b is performed. For example, when the output value matches the original data, the control unit 16 determines that the performance of the feature extraction layer 11b satisfies a certain standard. On the other hand, when the output value does not match the original data, the control unit 16 determines that the performance of the feature extraction layer 11b does not satisfy a certain standard.

(Parameter update of feature extraction layer)
Next, as shown in FIG. 4, the control unit 16 updates the parameters of the feature extraction layer 11b based on the result of the performance evaluation E1 in S3 (S4 to S6). Specifically, when the result of the performance evaluation E1 in S3 satisfies a certain criterion (Yes in S4), the parameter of the feature extraction layer 11b is updated (S5). For example, the parameter of the feature extraction layer 11b (for example, the weight w1) is updated with the parameter (for example, the weight w1') obtained by the additional learning L1. On the other hand, when the result of the performance evaluation E1 in S3 does not satisfy a certain standard (No in S4), the control unit 16 sets the parameter of the feature extraction layer 11b to the parameter before the additional learning L1 (No). For example, the weight is maintained at w1) (S6).

<Second learning / evaluation / update process>
The second learning / evaluation / updating step (III) includes a step of performing additional learning L2 of the identification layer 11c (S7), a step of performing performance evaluation E2 of the identification layer 11c by the additional learning L2 (S8), and performance evaluation. The step (S9 to S11) of updating the parameter (weight) of the identification layer 11c to the parameter obtained by the additional learning L2 based on the result of E2 is included.

(Additional learning L2)
In S7, after the first learning / evaluation / updating step (II), the additional learning L2 of the identification layer 11c is performed using the data set A with the correct answer. FIG. 8 schematically shows a data propagation model during the additional learning L2 of the discrimination layer 11c. In S7, the discrimination layer 11c'used in S2 to S6 is returned to the discrimination layer 11c used in the pre-learning of S1, and additional learning L2 of the discrimination layer 11c is performed by the same procedure as the pre-learning.

That is, the data of the data set A is input to the feature extraction layer 11b via the input layer 11a. In each unit U of the feature extraction layer 11b, a weighted operation is performed on the input data. By this calculation, the feature extraction layer 11b extracts the feature from the input data, and the data indicating the extracted feature is output to the identification layer 11c. In each unit U of the identification layer 11c, a weighted operation is performed on the input data, whereby identification based on the above characteristics is performed. The data indicating the identification result is output from the identification layer 11c.

The control unit 16 compares the output value (data) of the identification layer 11c with the value indicating the correct answer, calculates these errors (losses), and changes the weight of the identification layer 11c so that this error becomes small. (Backpropagation). As a result, the identification layer 11c is additionally learned. In the additional learning L2, the weights of not only the identification layer 11c but also the feature extraction layer 11b may be changed so as to reduce the above error, but here, the additional learning is limited to the identification layer 11c. (Additional learning of the feature extraction layer 11b is not performed). The reason will be described later.

(Performance evaluation E2)
In the step of S8, the control unit 16 performs the performance evaluation E2 of the identification layer 11c by the additional learning L2 of S7. More specifically, the control unit 16 performs the performance evaluation E2 of the identification layer 11c using the data set C. The data set C is a data set prepared in advance for performance evaluation, and is stored in the learning / evaluation data storage unit 12b.

FIG. 9 schematically shows a data propagation model at the time of performance evaluation E2 of the identification layer 11c. In S8, the data of the evaluation data set C is input to the input layer 11a. The above data input to the input layer 11a is output to the feature extraction layer 11b. In the feature extraction layer 11b, a weighted operation is performed on the input data, and the data indicating the operation result is output to the identification layer 11c. In the identification layer 11c, a weighted operation is performed on the data input from the feature extraction layer 11b, and the data indicating the calculation result is output from the identification layer 11c.

The control unit 16 performs the performance evaluation E2 of the identification layer 11c by determining whether or not the performance value (for example, the correct answer rate) obtained based on the output value of the identification layer 11c is equal to or higher than a certain threshold value. For example, when the performance value is equal to or higher than the threshold value, the control unit 16 determines that the performance of the identification layer 11b satisfies a certain standard. On the other hand, when the performance value is less than the threshold value, the control unit 16 determines that the performance of the identification layer 11c does not satisfy a certain standard.

(Parameter update of identification layer)
Next, the control unit 16 updates the parameters of the identification layer 11c based on the result of the performance evaluation E2 in S8 (S9 to S11). Specifically, when the result of the performance evaluation E2 in S8 satisfies a certain criterion (Yes in S9), the control unit 16 updates the parameter of the identification layer 11c (S10). For example, the parameter of the identification layer 11c (for example, the weight w2) is updated with the parameter (for example, the weight w2') obtained by the additional learning L2. On the other hand, when the result of the performance evaluation E2 in S8 does not satisfy a certain standard (No in S9), the control unit 16 sets the parameter of the identification layer 11c to the parameter before the additional learning L2 (for example,). The weight is maintained at w2) (S11).

In the second learning / evaluation / updating step (III) described above, the additional learning L2 is performed only in the identification layer 11c of the recognition processing unit 11, but the feature extraction layer 11b is subjected to backpropagation. Additional learning may be performed. Then, in the parameter update based on the performance evaluation E2, the parameter may be updated not only for the identification layer 11c but also for the feature extraction layer 11b. That is, in the second learning / evaluation / updating step (III), the target of the additional learning L2 and the parameter update may be the recognition processing unit 11 including the feature extraction layer 11b and the identification layer 11c.

However, the feature extraction layer 11b has already been additionally learned so as to be compatible not only with the data set A but also with the data set B in the first learning / evaluation / updating step (II). Therefore, in the second learning / evaluation / updating step (III), when the additional learning L2 based on the data set A is performed up to the feature extraction layer 11b, the compatibility of the feature extraction layer 11b with the data set B becomes the data set A. There is a concern that it will be pulled down to fit into. Therefore, from the viewpoint of improving the compatibility with the data acquired from the outside, it is desirable that the additional learning L2 in the second learning / evaluation / updating step (III) is performed only on the identification layer 11c.

The first learning / evaluation / updating step (II) and the second learning / evaluation / updating step (III) described above are performed in parallel with the acquisition of the data set B (while acquiring the data set B at the site). They may be performed at the same time (at the same time), or at different timings (for example, after the store is closed) than when the data set B is acquired.

FIG. 10 shows, using the learning device 1 learned by the above learning methods (steps (I) to (III)), actually recognizing (for example, customer behavior recognition) by inputting data obtained at a site (for example, a store). ) Is shown. When the learning device 1 is used (during operation), the feature extraction layer 11b whose parameters are updated according to the result of the performance evaluation E1 and the identification layer 11c (learning) whose parameters are updated according to the result of the performance evaluation E2. Feature extraction and identification will be performed using (not the dedicated identification layer 11c').

〔effect〕
As described above, the learning method of the present embodiment includes a pre-learning step (1) in which pre-learning of the recognition processing unit 11 is performed using a data set A with a correct answer obtained in advance, and data after pre-learning. The additional learning step (2) for performing additional learning of the recognition processing unit 11 using the set A and the data set B acquired in an environment different from the data set A, and the performance of the recognition processing unit 11 by the additional learning. It includes a performance evaluation step (3) for evaluation and an update step (4) for updating the parameters of the recognition processing unit 11 to the parameters obtained by the additional learning based on the result of the performance evaluation. The pre-learning step (1) corresponds to the above-mentioned S1 process, the additional learning step (2) corresponds to each of the above-mentioned S2 and S7 processes, and the performance evaluation step (3) corresponds to the above-mentioned S3. And each step of S8, and the update step (4) corresponds to each step of S4 to S6 and S9 to S11 described above.

Further, the learning device 1 of the present embodiment has a recognition processing unit 11 that extracts features from input data and performs identification using the extracted features, and a control unit 16 as a learning control unit that causes the recognition processing unit 11 to perform machine learning. The control unit 16 is acquired in a different environment from the data set A and the data set A for the recognition processing unit 11 that has been pre-learned using the data set A with the correct answer obtained in advance. After additional learning using the data set B, the performance of the recognition processing unit 11 is evaluated by the additional learning, and the parameters of the recognition processing unit 11 are obtained by the additional learning based on the result of the performance evaluation. Update to the specified parameters.

According to the above learning method and the configuration of the learning device 1, after the pre-learning of the recognition processing unit 11 is performed using the data set A with the correct answer, additional learning is performed using the data set A and the data set B. .. As a result, not only the data set A but also the data set B (for example, the data obtained at the site where the recognition processing unit 11 is actually desired to be used) obtained in an environment different from the data set A is recognized. It is possible to improve the compatibility of the processing unit 11 and improve the generalization.

Moreover, after performing additional learning, the performance evaluation (verification) of the recognition processing unit 11 is actually performed, and the parameters of the recognition processing unit 11 are updated based on the result. As a result, for example, the parameter is updated only when the performance of the recognition processing unit 11 is actually improved by the additional learning, while the parameter is updated when the performance of the recognition processing unit 11 is deteriorated by the additional learning (for example, overfitting). It is possible to keep the parameters before additional learning as they are without performing additional learning. Therefore, even when pre-learning and additional learning are performed, it is possible to avoid deterioration in performance (due to overfitting) of the recognition processing unit 11 and improve generalization.

In the above learning method, the additional learning step (2) is a step (2a) of performing the additional learning L1 of the feature extraction layer 11b and a step (2b) of performing the additional learning L2 of the recognition processing unit 11 (for example, the identification layer 11c). And include. The step (2a) corresponds to the step of S2, and the step (2b) corresponds to the step of S7. Further, the performance evaluation step (3) includes a step (3a) of performing the performance evaluation E1 of the feature extraction layer 11b by the additional learning L1 and a step (3b) of performing the performance evaluation E2 of the recognition processing unit 11 by the additional learning L2. include. The step (3a) corresponds to the step of S3, and the step (3b) corresponds to the step of S8. Further, the update step (4) is based on the step (4a) of updating the parameters of the feature extraction layer 11b to the parameters obtained by the additional learning L1 based on the results of the performance evaluation E1 and the results of the performance evaluation E2. The step (4b) of updating the parameters of the recognition processing unit 11 (for example, the identification layer 11c) to the parameters obtained by the additional learning L2 is included. The step (4a) corresponds to the steps of S4 to S6, and the step (4b) corresponds to the steps of S9 to S11. Then, in this embodiment, the step (2a), the step (3a), the step (4a), the step (2b), the step (3b), and the step (4b) are performed in this order.

Further, in the above learning device 1, the control unit 16 performs additional learning L1 on the feature extraction layer 11b, and then performs performance evaluation E1 of the feature extraction layer 11b by the additional learning L1, and the result of the performance evaluation E1 is obtained. Based on this, the parameters of the feature extraction layer 11b are updated to the parameters obtained by the additional learning L1, and subsequently, the additional learning L2 of the recognition processing unit 11 (for example, the identification layer 11c) is performed, and then the recognition by the additional learning L2. The performance evaluation E2 of the processing unit 11 is performed, and the parameters of the recognition processing unit 11 are updated to the parameters obtained by the additional learning L2 based on the result of the performance evaluation E2.

When the recognition processing unit 11 includes the feature extraction layer 11b and the identification layer 11c, the recognition processing unit including the identification layer 11c is first subjected to additional learning L1, performance evaluation E1 and parameter update of the feature extraction layer 11b. 11 additional learning L2, performance evaluation E2, and parameter update are performed. As a result, during the actual operation of the recognition processing unit 11 (when the recognition processing is actually performed using the field data), the identification layer 11c uses the features appropriately extracted from the input data by the feature extraction layer 11b. Can be properly identified at.

In the above learning method, in the step (2a), the additional learning L1 of the feature extraction layer 11b is performed using the data set B1 included in the data set B, and in the step (3a), the data set included in the data set B is performed. Performance evaluation E1 of the feature extraction layer 11b is performed using a data set B2 other than B1. Further, in the above learning device 1, the control unit 16 performs additional learning L1 of the feature extraction layer 11b using the data set B1 included in the data set B, while the control unit 16 other than the data set B1 included in the data set B. The performance evaluation E1 of the feature extraction layer 11b is performed using the data set B2.

The additional learning L1 of the feature extraction layer 11b is performed using the data set B1 included in the data set B, and the performance evaluation E1 of the feature extraction layer 11b is performed by the data set B2 other than the data set B1 included in the data set B. Whether or not the feature extraction layer 11b exhibits appropriate performance even when data other than the data used in the additional learning L1 is input (whether or not the features are properly extracted from the input data). Can be evaluated. This also leads to improved generalization.

In the above learning method, in the step (2b), the additional learning L2 of the recognition processing unit 11 (for example, the identification layer 11c) is performed using the data set A, and in the step (3b), the data for performance evaluation prepared in advance is performed. The performance evaluation E2 of the recognition processing unit 11 is performed using the set C. Further, in the above learning device 1, the control unit 16 performs additional learning L2 of the recognition processing unit 11 (for example, the identification layer 11c) using the data set A, while the data set C for performance evaluation prepared in advance is used. The performance evaluation E2 of the recognition processing unit 11 is performed using the data.

By performing the additional learning L2 of the recognition processing unit 11 using the data set A with a correct answer, the recognition processing unit 11 can be trained so that the output from the recognition processing unit 11 approaches the correct answer. Further, by performing the performance evaluation E2 of the recognition processing unit 11 using the data set C for performance evaluation prepared in advance, the performance evaluation of the recognition processing unit 11 can be appropriately performed.

In the above learning method, in the step (4a), when the result of the performance evaluation E1 satisfies a certain criterion, the parameter of the feature extraction layer 11b is updated (S5). Further, in the learning device 1 described above, the control unit 16 updates the parameters of the feature extraction layer 11b when the result of the performance evaluation E1 satisfies a certain criterion.

When the result of the performance evaluation E1 satisfies a certain criterion, the performance of the feature extraction layer 11b can be improved as compared with that before the additional learning L1 by updating the parameters of the feature extraction layer 11b.

In the above learning method, in the step (4a), when the result of the performance evaluation E1 does not satisfy a certain criterion, the parameter of the feature extraction layer 11b is maintained at the parameter before the additional learning L1 (S6). ). Further, in the learning device 1 described above, the control unit 16 maintains the parameters of the feature extraction layer 11b at the parameters before the additional learning L1 when the result of the performance evaluation E1 does not satisfy a certain standard. ..

When the result of the performance evaluation E1 does not satisfy a certain standard, the parameter of the feature extraction layer 11b is maintained at the parameter before the additional learning L1, so that the performance of the feature extraction layer 11b is additionally learned. It can be avoided to decrease due to L1 (for example, overfitting).

In the above learning method, in the step (4b), when the result of the performance evaluation E2 satisfies a certain criterion, the parameter of the recognition processing unit 11 (for example, the identification layer 11c) is updated (S10). Further, in the learning device 1 described above, the control unit 16 updates the parameters of the recognition processing unit (for example, the identification layer 11c) when the result of the performance evaluation E2 satisfies a certain standard.

When the result of the performance evaluation E2 satisfies a certain standard, the performance of the recognition processing unit 11 can be improved as compared with that before the additional learning L2 by updating the parameters of the recognition processing unit 11.

In the above learning method, in the step (4b), when the result of the performance evaluation E2 does not satisfy a certain criterion, the parameters of the recognition processing unit 11 (for example, the identification layer 11c) are set before the additional learning L2 is performed. Maintained as a parameter (S11). Further, in the above learning device 1, when the result of the performance evaluation E2 does not satisfy a certain standard, the control unit 16 performs additional learning L2 on the parameters of the recognition processing unit 11 (for example, the identification layer 11c). Keep the previous parameters.

When the result of the performance evaluation E2 does not satisfy a certain standard, the parameter of the recognition processing unit 11 is maintained at the parameter before the additional learning L2, so that the performance of the recognition processing unit 11 is subjected to the additional learning. It can be avoided to decrease due to L2 (for example, overfitting).

In the above learning method, in the step (2b), the additional learning L2 using the data set A is performed only for the identification layer 11c. Further, in the learning device 1 described above, the control unit 16 performs additional learning L2 using the data set A only on the identification layer 11c.

By performing the additional learning L2 only on the identification layer 11c, the feature extraction layer 11b that was previously appropriately tuned (parameter adjusted) using the data set B by the additional learning L1 is pulled to the data set A with a correct answer. You can prevent it from being learned. This makes it possible to appropriately learn only the identification layer 11c while preventing overfitting of the feature extraction layer 11b and preventing performance deterioration of the feature extraction layer 11b.

In the above learning method and learning device 1, the recognition processing unit 11 is composed of a neural network. In this case, the performance (generalization) of the recognition processing unit 11 can be improved by machine learning of the neural network.

[Programs and recording media]
The learning device 1 described in this embodiment can be configured by, for example, a computer (PC) in which a predetermined program (application software) is installed. By reading and executing the above program by a computer (for example, a control unit 16 as a CPU), each unit of the learning device 1 can be operated to execute each process (each step) described above. Such a program is acquired by downloading from the outside via a network, for example, and is stored in the program storage unit 12a. Further, the program is recorded on a computer-readable recording medium such as a CD-ROM (Compact Disk-Read Only Memory), and the computer reads the program from the recording medium and stores it in the program storage unit 12a. There may be.

〔supplement〕
FIG. 11 is a flowchart showing the flow of other processes of the learning method of the present embodiment. The control unit 16 determines whether or not the result of the performance evaluation E1 in S3 satisfies a certain standard (S4), and if it does not satisfy a certain standard, returns to S2 and has a correct answer. The additional learning L1 of the feature extraction layer 11b may be performed again using the data set A and the data set B3 other than the data set B1 included in the data set B. When the result of the performance evaluation E1 does not satisfy a certain standard, the performance improvement of the feature extraction layer 11b can be tried again by performing the additional learning L1 of the feature extraction layer 11b again, thereby making the feature extraction layer 11b generalizable. Improvement can be expected.

In the present embodiment, an example in which the recognition processing unit 11 is configured by a neural network has been described, but the present invention is not limited to this configuration. For example, a support vector machine (SVM), a random forest, or the like can be used as the identification layer 11c of the recognition processing unit 11, and even in this case, the learning method described in the present embodiment is applied. It is possible to do.

〔summary〕
In the present embodiment, the learning results by the additional learning L1 and L2 of the recognition processing unit 11 are evaluated using the evaluation data, and the parameters of each layer (feature extraction layer 11b, identification layer 11c) are updated based on the evaluation results. I do. If it is expected that the accuracy of feature extraction or identification will be improved based on the evaluation of the learning result, the parameters of the feature extraction layer 11b and the identification layer 11c will be updated respectively, so that the accuracy of feature extraction or identification will be improved and generalized. It is possible to improve the creativity.

Further, in performing machine learning of the recognition processing unit 11 including the feature extraction layer 11b and the identification layer 11c, the actual recognition (identification) is actually performed only by learning using the data with the correct answer (data set A) collected in advance. ) Is poorly compatible with the local data (data set B), and generalization cannot be improved. In addition, the local data cannot be used for learning the discrimination layer 11c because there is no annotation even if it is collected and it is not known whether the output of the discrimination layer 11c is the correct answer. However, by training the recognition processing unit 11 using the local data together with the data with the correct answer collected in advance, the data to be used is diversified, and learning according to the local data becomes possible. Increases versatility. That is, as in the present embodiment, the data with the correct answer (data set A) obtained in advance and the data (data set B) acquired at the site where the actual operation is performed are used to combine the data and the learning range. By performing pre-learning and additional learning L1 and L2 while adjusting the above, generalization can be improved. Further, the accuracy of identification can be improved by performing additional learning of the identification layer 11c using the data with the correct answer.

In particular, the additional learning L1 of the feature extraction layer 11b is performed by using the data set A and the data set B to improve generalization, and the additional learning L2 of the identification layer 11c is performed by using the data set A with a correct answer. By doing so, the accuracy of identification can be improved. Further, by not performing the additional learning (parameter learning) of the feature extraction layer 11b at the time of the additional learning L2 of the identification layer 11c, it is possible to suppress over-learning of the feature extraction layer 11b to the data set A. ..

〔others〕
The learning method and learning device of the present embodiment described above may be expressed as follows. In addition, the contents described in this embodiment include the following programs and recording media.

1. 1. It is a learning method in machine learning of the recognition processing unit that extracts features from input data and identifies using the extracted features.
The pre-learning step (1) in which the recognition processing unit is pre-learned using the data set A with the correct answer obtained in advance, and
After the pre-learning, an additional learning step (2) of performing additional learning of the recognition processing unit using the data set A and the data set B without a correct answer acquired in an environment different from the data set A. )When,
The performance evaluation step (3) for evaluating the performance of the recognition processing unit by the additional learning, and
A learning method comprising an update step (4) for updating a parameter of the recognition processing unit to a parameter obtained by the additional learning based on the result of the performance evaluation.

2. 2. When the recognition processing unit includes a feature extraction layer that extracts the feature from the input data and an identification layer that performs identification using the extracted feature.
The additional learning step (2) includes a step (2a) of performing the additional learning L1 of the feature extraction layer and a step (2b) of performing the additional learning L2 of the recognition processing unit.
The performance evaluation step (3) includes a step (3a) of performing the performance evaluation E1 of the feature extraction layer by the additional learning L1 and a step (3b) of performing the performance evaluation E2 of the recognition processing unit by the additional learning L2. Including
The update step (4) is a step (4a) of updating the parameters of the feature extraction layer to the parameters obtained by the additional learning L1 based on the result of the performance evaluation E1, and the result of the performance evaluation E2. The step (4b) of updating the parameters of the recognition processing unit to the parameters obtained by the additional learning L2 is included.
2. The above 1 is characterized in that the step (2a), the step (3a), the step (4a), the step (2b), the step (3b), and the step (4b) are performed in this order. Learning method.

3. 3. In the step (2a), the additional learning L1 of the feature extraction layer is performed using the data set B1 included in the data set B.
2. The learning method according to 2, wherein in the step (3a), the performance evaluation E1 of the feature extraction layer is performed using a data set B2 other than the data set B1 included in the data set B. ..

4. In the step (2b), the additional learning L2 of the recognition processing unit is performed using the data set A.
The learning method according to 2 or 3, wherein in the step (3b), the performance evaluation E2 of the recognition processing unit is performed using the data set C for performance evaluation prepared in advance.

5. The learning according to any one of 2 to 4, wherein in the step (4a), when the result of the performance evaluation E1 satisfies a certain criterion, the parameter of the feature extraction layer is updated. Method.

6. The step (4a) is characterized in that, when the result of the performance evaluation E1 does not satisfy a certain criterion, the parameter of the feature extraction layer is maintained at the parameter before the additional learning L1 is performed. The learning method according to any one of 2 to 5 above.

7. The learning according to any one of 2 to 6 above, wherein in the step (4b), when the result of the performance evaluation E2 satisfies a certain criterion, the parameter of the recognition processing unit is updated. Method.

8. The step (4b) is characterized in that, when the result of the performance evaluation E2 does not satisfy a certain criterion, the parameter of the recognition processing unit is maintained at the parameter before the additional learning L2 is performed. The learning method according to any one of 2 to 7 above.

9. The learning method according to any one of 2 to 8, wherein in the step (2b), the additional learning L2 using the data set A is performed only on the identification layer.

10. The learning method according to any one of 1 to 9, wherein the recognition processing unit is composed of a neural network.

11. A recognition processing unit that extracts features from input data and uses the extracted features for identification.
The recognition processing unit is provided with a learning control unit for machine learning.
The learning control unit was acquired in a different environment from the data set A and the data set A for the recognition processing unit that had been pre-learned using the data set A with the correct answer obtained in advance. After additional learning using the data set B without a correct answer, the performance evaluation of the recognition processing unit is performed by the additional learning, and the parameters of the recognition processing unit are subjected to the additional learning based on the result of the performance evaluation. A learning device characterized by updating to the parameters obtained by.

12. The recognition processing unit has a feature extraction layer that extracts the feature from the input data, and an identification layer that performs identification using the extracted feature.
The learning control unit
After the additional learning L1 is performed on the feature extraction layer, the performance evaluation E1 of the feature extraction layer is performed by the additional learning L1, and the parameters of the feature extraction layer are set based on the result of the performance evaluation E1. After updating to the parameters obtained by the additional learning L1 and subsequently performing the additional learning L2 of the recognition processing unit, the performance evaluation E2 of the recognition processing unit by the additional learning L2 is performed, and the result of the performance evaluation E2. 11. The learning apparatus according to 11 above, wherein the parameters of the recognition processing unit are updated to the parameters obtained by the additional learning L2.

13. The learning control unit performs the additional learning L1 of the feature extraction layer using the data set B1 included in the data set B, while the data set B2 other than the data set B1 included in the data set B is used. The learning apparatus according to 12 above, wherein the performance evaluation E1 of the feature extraction layer is performed by using the learning apparatus.

14. The learning control unit performs the additional learning L2 of the recognition processing unit using the data set A, while performing the performance evaluation E2 of the recognition processing unit using the performance evaluation data set C prepared in advance. 12. The learning device according to 12 or 13, wherein the learning device is performed.

15. The learning device according to any one of 12 to 14, wherein the learning control unit updates the parameters of the feature extraction layer when the result of the performance evaluation E1 satisfies a certain criterion. ..

16. The learning control unit is characterized in that when the result of the performance evaluation E1 does not satisfy a certain criterion, the parameter of the feature extraction layer is maintained at the parameter before the additional learning L1 is performed. The learning device according to any one of 12 to 15.

17. The learning device according to any one of 12 to 16, wherein the learning control unit updates the parameters of the recognition processing unit when the result of the performance evaluation E2 satisfies a certain criterion. ..

18. The learning control unit is characterized in that when the result of the performance evaluation E2 does not satisfy a certain criterion, the parameter of the recognition processing unit is maintained at the parameter before the additional learning L2 is performed. The learning device according to any one of 12 to 17.

19. The learning device according to any one of 12 to 18, wherein the learning control unit performs the additional learning L2 using the data set A only for the identification layer.

20. The learning device according to any one of 11 to 19, wherein the recognition processing unit is composed of a neural network.

21. In the step (2a), when the result of the performance evaluation E1 does not satisfy a certain criterion in the step (4a), the data set A and the unused data set B3 included in the data set B are used. The learning method according to any one of 2 to 5, wherein the additional learning L1 of the feature extraction layer is performed again using the above.

22. When the result of the performance evaluation E1 does not satisfy a certain criterion, the learning control unit uses the data set A and the unused data set B3 included in the data set B to perform the feature. The learning apparatus according to any one of 12 to 15, wherein the additional learning L1 of the extraction layer is performed again.

23. A program that causes a computer to execute the learning method according to any one of 1 to 10 and 21.

24. A computer-readable recording medium on which the program according to 23 is recorded.

Although the embodiments of the present invention have been described above, the scope of the present invention is not limited to this, and can be extended or modified without departing from the gist of the invention.

INDUSTRIAL APPLICABILITY The present invention can be used as a learning method and a learning device in machine learning of a recognition processing unit that extracts features from input data and identifies using the extracted features.

1 Learning device 11 Recognition processing unit 11b Feature extraction layer 11c Identification layer 16 Control unit (learning control unit)

Claims (20)

It is a learning method in machine learning of the recognition processing unit that extracts features from input data and identifies using the extracted features.
The pre-learning step (1) in which the recognition processing unit is pre-learned using the data set A with the correct answer obtained in advance, and
After the pre-learning, an additional learning step (2) of performing additional learning of the recognition processing unit using the data set A and the data set B without a correct answer acquired in an environment different from the data set A. )When,
The performance evaluation step (3) for evaluating the performance of the recognition processing unit by the additional learning, and
A step (4) of updating the parameters of the recognition processing unit to the parameters obtained by the additional learning based on the result of the performance evaluation is included.
When the recognition processing unit includes a feature extraction layer that extracts the feature from the input data and an identification layer that performs identification using the extracted feature.
The additional learning step (2) includes a step (2a) of performing the additional learning L1 of the feature extraction layer and a step (2b) of performing the additional learning L2 of the recognition processing unit.
The performance evaluation step (3) includes a step (3a) of performing the performance evaluation E1 of the feature extraction layer by the additional learning L1 and a step (3b) of performing the performance evaluation E2 of the recognition processing unit by the additional learning L2. Including
The update step (4) is a step (4a) of updating the parameters of the feature extraction layer to the parameters obtained by the additional learning L1 based on the result of the performance evaluation E1, and the result of the performance evaluation E2. The step (4b) of updating the parameters of the recognition processing unit to the parameters obtained by the additional learning L2 is included.
A learning method characterized in that the step (2a), the step (3a), the step (4a), the step (2b), the step (3b), and the step (4b) are performed in this order. In the step (2a), the additional learning L1 of the feature extraction layer is performed using the data set B1 included in the data set B.
The learning according to claim 1, wherein in the step (3a), the performance evaluation E1 of the feature extraction layer is performed using a data set B2 other than the data set B1 included in the data set B. Method. In the step (2b), the additional learning L2 of the recognition processing unit is performed using the data set A.
The learning method according to claim 1 or 2, wherein in the step (3b), the performance evaluation E2 of the recognition processing unit is performed using the data set C for performance evaluation prepared in advance. The method according to any one of claims 1 to 3, wherein in the step (4a), the parameter of the feature extraction layer is updated when the result of the performance evaluation E1 satisfies a certain criterion. Learning method. The step (4a) is characterized in that, when the result of the performance evaluation E1 does not satisfy a certain criterion, the parameter of the feature extraction layer is maintained at the parameter before the additional learning L1 is performed. The learning method according to any one of claims 1 to 4. The method according to any one of claims 1 to 5, wherein in the step (4b), when the result of the performance evaluation E2 satisfies a certain criterion, the parameter of the recognition processing unit is updated. Learning method. The step (4b) is characterized in that, when the result of the performance evaluation E2 does not satisfy a certain criterion, the parameter of the recognition processing unit is maintained at the parameter before the additional learning L2 is performed. The learning method according to any one of claims 1 to 6. The learning method according to any one of claims 1 to 7, wherein in the step (2b), the additional learning L2 using the data set A is performed only on the identification layer. The learning method according to any one of claims 1 to 8, wherein the recognition processing unit is composed of a neural network. A program that causes a computer to execute the learning method according to any one of claims 1 to 9. A computer-readable recording medium on which the program according to claim 10 is recorded. A recognition processing unit that extracts features from input data and uses the extracted features for identification.
The recognition processing unit is provided with a learning control unit for machine learning.
The learning control unit was acquired in a different environment from the data set A and the data set A for the recognition processing unit that had been pre-learned using the data set A with the correct answer obtained in advance. After additional learning using the data set B without a correct answer, the performance of the recognition processing unit is evaluated by the additional learning, and the parameters of the recognition processing unit are subjected to the additional learning based on the result of the performance evaluation. Update to the parameters obtained by
The recognition processing unit has a feature extraction layer that extracts the feature from the input data, and an identification layer that performs identification using the extracted feature.
The learning control unit
After performing the additional learning L1 on the feature extraction layer, the performance evaluation E1 of the feature extraction layer is performed by the additional learning L1, and the parameters of the feature extraction layer are set based on the result of the performance evaluation E1. After updating to the parameters obtained by the additional learning L1 and subsequently performing the additional learning L2 of the recognition processing unit, the performance evaluation E2 of the recognition processing unit by the additional learning L2 is performed, and the result of the performance evaluation E2. A learning device characterized in that the parameters of the recognition processing unit are updated to the parameters obtained by the additional learning L2 based on the above. The learning control unit performs the additional learning L1 of the feature extraction layer using the data set B1 included in the data set B, while the data set B2 other than the data set B1 included in the data set B is used. The learning device according to claim 12, wherein the performance evaluation E1 of the feature extraction layer is performed by using the learning device. The learning control unit performs the additional learning L2 of the recognition processing unit using the data set A, while performing the performance evaluation E2 of the recognition processing unit using the performance evaluation data set C prepared in advance. The learning device according to claim 12 or 13, wherein the learning device is performed. The learning according to any one of claims 12 to 14, wherein the learning control unit updates the parameters of the feature extraction layer when the result of the performance evaluation E1 satisfies a certain criterion. Device. The claim is characterized in that, when the result of the performance evaluation E1 does not satisfy a certain criterion, the learning control unit maintains the parameter of the feature extraction layer to the parameter before the additional learning L1 is performed. Item 6. The learning device according to any one of Items 12 to 15. The learning according to any one of claims 12 to 16, wherein the learning control unit updates the parameters of the recognition processing unit when the result of the performance evaluation E2 satisfies a certain criterion. Device. The claim is characterized in that, when the result of the performance evaluation E2 does not satisfy a certain criterion, the learning control unit maintains the parameters of the recognition processing unit to the parameters before the additional learning L2 is performed. Item 6. The learning device according to any one of Items 12 to 17. The learning device according to any one of claims 12 to 18, wherein the learning control unit performs the additional learning L2 using the data set A only for the identification layer. The learning device according to any one of claims 12 to 19, wherein the recognition processing unit is composed of a neural network.

Images