JP6516531B2 - Clustering device and machine learning device - Google Patents

Description

  The present invention relates to a clustering apparatus and a machine learning apparatus used in machine learning in which transfer learning is introduced.

  Machine learning is used for processing of detecting a person from image data and analysis of measurement data by a sensor.

  For example, when detecting a person from an image captured by a surveillance camera, identification feature data generated by learning the characteristics of the person is used. Specifically, the machine learning apparatus learns the characteristics of a person using a plurality of images (a plurality of learning samples) in which the person is photographed, and generates identification feature data reflecting the learning result. The person detection device detects a person from the image captured by the surveillance camera using the identification feature data generated by the machine learning device.

  When the installation environment of the surveillance camera is different from the environment in which the learning sample is collected, the appearance of the person photographed by the surveillance camera is different from the appearance of the person in the learning sample. That is, the feature of the person photographed by the surveillance camera is different from the feature of the person included in the learning sample. Therefore, when the identification feature data generated from the learning sample is used to detect a person from the image generated by the surveillance camera, the detection accuracy of the person decreases. In order to improve the detection accuracy of a person, a huge number of learning samples must be prepared according to the installation environment of the camera, which increases the cost.

  Therefore, a method of machine learning that introduces transfer learning has been proposed. Transfer learning is a method of learning in advance a sample obtained from an environment different from the collection environment of learning samples, and applying (transferring) the feature of the detection target obtained by the prior learning to the learning result of the learning sample. Since transfer learning can suppress the number of learning samples, the cost for generating identification feature data can be reduced.

  In transfer learning, a set of samples that are learned in advance is called a prior domain. The target to which the learning result of the prior domain is transferred is called the target domain. When detecting a person from an image captured by a surveillance camera, the target domain is a set of learning samples generated in accordance with the installation environment of the surveillance camera. The pre-domain is a set of learning samples generated in an environment different from the installation environment of the surveillance camera.

Xiaoxiao Shi, Wei Fan, Jiangtao Ren, "Actively Transfer Domain Knowledge", [online], [search on March 1, 2015], Internet <URL: www.cs.columbia.edu/~wfan/PAPERS/ecml08transfer. pdf>

  It is known that, when using transfer learning, a phenomenon called negative transfer occurs. Negative transfer is a phenomenon in which the accuracy of learning decreases when the pre-learned in advance for transfer learning includes data significantly different from the data included in the target domain. For this reason, before performing machine learning in which transfer learning is introduced, it is desirable to specify an advance domain effective for transfer learning, and use only the specified advance domain for machine learning.

  Non-Patent Document 1 discloses a method of determining whether a prior domain is effective for transfer learning. Specifically, in the method according to Non-Patent Document 1, a classifier (pre-classifier) learned using only the pre-domain, and a classifier that performed transition learning using the pre-domain and the target domain Input the sample data to the If the discrimination result by the pre-discriminator for the sample data is the same as the discrimination result by the metastasis discriminator, this pre-domain is judged to be effective for metastasis learning.

  As a result, in the method disclosed in Non-Patent Document 1, the pre-domain determined not to be effective for transfer learning is not used for machine learning in which transfer learning is introduced. If the number of pre-domains to be introduced into transfer learning is one and it is determined that this pre-domain is not effective for transfer learning, machine learning with transfer learning can not be performed.

  Therefore, when determining whether a preliminary domain is effective for transfer learning, it is desirable to prepare a plurality of preliminary domains in advance. However, a method in which a human confirms the collected samples one by one and classifies a plurality of prior domains is not realistic. Also, no technology has been developed to efficiently create multiple pre-domains from collected data.

  An object of the present invention is to provide a technique for efficiently creating a plurality of preliminary domains from a plurality of data collected to create preliminary domains in view of the above-mentioned problems.

  In order to solve the above problems, the invention according to claim 1 is a clustering device, and a feature is extracted from each of a plurality of transfer candidate data used for machine learning in which transfer learning is introduced, and a plurality of transfer candidate feature data Each transition candidate feature data includes a first group and a second group based on the features of each of the plurality of transition candidate feature data generated by the clustering feature extraction unit for generating The first group is used for the machine learning when the classification unit for classifying into a plurality of groups and the number of transfer candidate feature data classified into the first group by the classification unit is equal to or less than a predetermined classification continuation reference value Prior to the first group, and if the number of transfer candidate feature data is greater than the classification continuation reference value, the first group is classified. It comprises a pre-domain determining unit that determines that further classify transition candidate feature data, the.

  The invention according to claim 2 is the clustering apparatus according to claim 1, wherein the prior domain determination unit determines that the number of transfer candidate feature data classified into the first group is smaller than a predetermined discarding reference value. , Excluding the first group from the pre-domain.

  The invention according to claim 3 is the clustering apparatus according to claim 1 or 2, wherein the first group is further based on the feature value of each of the transfer candidate feature data classified into the first group. Calculating the variance of the transfer candidate feature data classified into two, wherein the prior domain determination unit determines that the number of transfer candidate feature data classified into the first group is higher than the classification continuation reference value If larger, the variance calculated by the variance calculator is compared with a predetermined variance reference value, and if the variance calculated by the variance calculator is less than or equal to the variance reference value, the first group is determined to be a pre-domain Do.

  The invention according to claim 4 is the clustering apparatus according to any one of claims 1 to 3, wherein the classification unit is configured to change the number of transfer candidate feature data classified into the first group to a predetermined change reference value If larger, the transition candidate feature data classified into the first group is further classified into a first number of lower groups, and the classification unit determines that the number of the transition candidate feature data classified into the first group is If it is equal to or less than the change reference value, transfer candidate feature data classified into the first group is classified into a second number of lower groups smaller than the first number.

  The invention according to claim 5 is the clustering apparatus according to any one of claims 1 to 5, wherein the classification continuation reference value corresponds to the number of dimensions of transfer candidate feature data extracted by the clustering feature extraction unit. It is decided based on.

  The invention according to claim 6 is a machine learning device that executes machine learning in which transfer learning is introduced to learn a detection target, and a plurality of transfer candidate data used for the machine learning are classified and the machine learning is performed. The clustering apparatus includes: a clustering apparatus that generates a pre-domain to be used; and a pre-domain evaluation apparatus that evaluates whether or not the pre-domain generated by the clustering apparatus is effective for the machine learning; A feature extraction unit for clustering which extracts features from each of the transfer candidate data of each to generate a plurality of transfer candidate feature data, and a feature possessed by each of a plurality of transfer candidate feature data generated by the clustering feature extraction unit Classification based on each transition candidate feature data into a plurality of groups including a first group and a second group And, if the number of transfer candidate feature data classified into the first group by the classification unit is equal to or less than a predetermined classification continuation reference value, the first group is determined to be a pre-domain used for the machine learning, A pre-domain determination unit that determines to further classify the transition candidate feature data classified into the first group if the number of metastasis candidate feature data is larger than the classification continuation reference value; The evaluation apparatus, when the first group is determined to be the prior domain by the prior domain determination unit, transfer candidate feature data included in the first group, and a feature to be detected under each of the predetermined conditions Perform the machine learning using a target domain including learning data having a to generate an evaluation discriminator for evaluating the pre-domain A row transfer learning unit, based on the trial transition identification unit generated by the trial transfer learning unit, and a determination portion that the first group to determine whether it is effective in the machine learning.

The invention according to claim 7 is the machine learning device according to claim 6, wherein the prior domain evaluation device further extracts a feature possessed by each of learning data included in the target domain, The trial transition learning unit executes the machine learning using the learning feature data, and the learning feature extraction unit is characterized by the learning data. The conditions for extracting are the same as the conditions for the feature extraction unit for clustering to extract features from each of the plurality of transfer candidate data.
The invention according to claim 8 is the machine learning device according to claim 7, further comprising: the target domain, and all the prior domains judged to be effective for the machine learning by the prior domain evaluation device. And a selection learning device that executes the machine learning to generate a transition identification unit,
Equipped with
The invention according to claim 9 is a clustering method, comprising the steps of: extracting features from each of a plurality of transfer candidate data used for machine learning in which transfer learning is introduced and generating a plurality of transfer candidate feature data; Classifying each of the transfer candidate feature data into a plurality of groups including a first group and a second group based on the features of each of the plurality of transfer candidate feature data, and the transfer classified into the first group When the number of candidate feature data is equal to or less than a predetermined classification continuation reference value, the first group is determined to be a pre-domain used for the machine learning, and the number of transfer candidate feature data is from the classification continuation reference value If it is also large, it is determined to further classify the transition candidate feature data classified into the first group.

  The invention according to claim 10 is a program for causing a computer to execute a clustering method for classifying each of a plurality of transfer candidate data used for machine learning into which transfer learning has been introduced, the plurality of used for the machine learning Extracting a feature from each of the transfer candidate data and generating a plurality of transfer candidate feature data; and based on the features possessed by each of the plurality of transfer candidate feature data generated, each transfer candidate feature data being a first group And classifying the first group into the plurality of groups including the second group, and when the number of transfer candidate feature data classified into the first group is equal to or less than a predetermined classification continuation reference value, Determining a prior domain to be used, and if the number of transfer candidate feature data is greater than the classification continuation reference value; Determining that the further classify transition candidate feature data classified into serial first group, is a program for executing a clustering method on a computer equipped with.

  In the machine learning apparatus of the present invention, the plurality of transfer candidate feature data generated from the plurality of transfer candidate data are classified into a plurality of groups including a first group and a second group based on the respective features. When the number of transfer candidate feature data classified into the first group is equal to or less than the classification continuation reference value, the first group is determined to be a pre-domain, and the number of transfer candidate feature data is larger than the classification continuation reference value The transfer candidate feature data classified into the first group is further classified. Thereby, it is possible to efficiently create the pre-domain used for machine learning in which transfer learning is introduced.

It is a functional block diagram showing composition of a machine learning device concerning an embodiment of the invention. It is a functional block diagram which shows the structure of the clustering apparatus shown in FIG. It is a functional block diagram which shows the structure of the prior domain evaluation apparatus shown in FIG. It is a functional block diagram which shows the structure of the selective learning apparatus shown in FIG. It is a flowchart which shows operation | movement of the machine learning apparatus shown in FIG. FIG. 7 is a view showing an example of distribution of transition candidate feature data generated from the transition candidate data shown in FIG. 1 and learning feature data generated from learning data. It is a figure which shows the range of the prior domain produced | generated by classify | categorizing the transfer candidate characteristic data shown in FIG. It is a flowchart of the prior domain production | generation process shown in FIG. FIG. 6 is a diagram showing an initial structure of a classification tree created in the pre-domain generation process shown in FIG. 5; It is a figure which shows an example of a structure when a node is added to the classification tree shown in FIG. It is a figure which shows an example of a structure of a classification tree when the prior domain production | generation process shown in FIG. 5 is complete | finished. It is a flowchart of the prior domain evaluation process shown in FIG. It is a figure which shows the modification of the classification | category tree shown in FIG.

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

[1. Configuration of Machine Learning Device 100]
[1.1. overall structure]
FIG. 1 is a functional block diagram showing the configuration of a machine learning apparatus 100 according to an embodiment of the present invention. The machine learning apparatus 100 shown in FIG. 1 executes machine learning in which transfer learning is introduced, using the plurality of transfer candidate data 41 stored in the storage device 400 and the target domain 50A stored in the storage device 500. . The machine learning apparatus 100 generates transition identification data 80 for identifying a detection target as a result of the above-described machine learning.

  In the present embodiment, the detection target is a person. The transition identification data 80 generated by the machine learning device 100 is used by a person detection device (not shown) to detect a person from an image captured by a camera. The machine learning apparatus 100 uses, as a learning algorithm for generating the transfer identification data 80, a random forest into which transfer learning is introduced. Therefore, the transition identification data 80 is a data group constituted by a plurality of decision trees.

  The storage device 500 stores the target domain 50A. The target domain 50A is a group of a plurality of images having features of a detection target (person) under predetermined conditions. The target domain 50A includes learning data 51, 51,. The learning data 51 is, for example, an image obtained by photographing a person at a depression angle of 0 °. The target domain 50A is used when the selection learning device 30 executes machine learning in which transfer learning is introduced to generate transfer identification data 80.

  The storage device 400 stores transfer candidate data 41, 41,. The plurality of pieces of transfer candidate data 41 are images in which a person is photographed, and are collected by searching the images in which the person is photographed on the Internet. By classifying the transfer candidate data 41, 41,... Based on the respective features of the transfer candidate data 41, 41,. Among the pre-domains 45, 45, ..., the pre-domain 45 judged to be effective for transfer learning is used for generation of transfer identification data 80.

  The machine learning device 100 includes a clustering device 10, a pre-domain evaluation device 20, and a selection learning device 30.

  The clustering device 10 classifies the transfer candidate data 41 based on the respective features of the transfer candidate data 41 to generate the pre-domain 45.

  The preliminary domain evaluation device 20 evaluates whether each of the preliminary domains 45 generated by the clustering device 10 is effective for transfer learning. The prior domain evaluation device 20 outputs evaluation result data 253A indicating the evaluation result of each prior domain 45 to the selective learning device 30.

  The selective learning device 30 selects, from among the pre-domains 45, the pre-domain 45 judged to be effective for transfer learning by the pre-domain evaluation device 20 based on the evaluation result data 253A. The selective learning device 30 executes machine learning in which transfer learning is introduced, using the selected prior domain 45 and the target domain 50A stored in the storage device 500. As a result, metastasis identification data 80 is generated.

[1.2. Configuration of Clustering Device 10]
FIG. 2 is a functional block diagram showing the configuration of the clustering device 10 shown in FIG. As shown in FIG. 2, the clustering device 10 includes a feature extraction unit 11, a classification unit 12, a distributed calculation unit 13, and a pre-domain determination unit 14.

  The clustering device 10 inputs a plurality of transfer candidate data 41 from the storage device 400. The feature extraction unit 11 extracts HOG (Histograms of Oriented Gradients) feature quantities from each of the plurality of transfer candidate data 41 input to the clustering apparatus 10, and a plurality of transfer candidate features corresponding to each of the transfer candidate data 41. Data 42 is generated. Hereinafter, the HOG feature amount is simply referred to as a “feature amount” unless otherwise described.

  The classification unit 12 receives the plurality of transfer candidate feature data 42 from the feature extraction unit 11. The classification unit 12 classifies the transition candidate feature data 42 into a plurality of groups based on the feature amounts included in each of the plurality of input transition candidate feature data 42. An algorithm called Density Forest is used to classify the transfer candidate feature data 42. The classification unit 12 classifies the plurality of transfer candidate feature data 42 while creating one classification tree. Each of the nodes constituting the classification tree corresponds to each group.

  The variance calculator 13 calculates the covariance of each node. The covariance of each node is calculated from the feature amount of the transfer candidate feature data 42 belonging to each node. The covariance of each node is used in classifying transition candidate feature data 42 belonging to each node. In addition, covariance is used to determine whether or not nodes constituting a classification tree are determined in advance.

  The prior domain determination unit 14 determines whether or not the nodes forming the classification tree satisfy the condition as the prior domain. If the number of transfer candidate feature data 42 belonging to the determination target node is equal to or less than the preset classification continuation reference value, the prior domain determination unit 14 determines the determination target node as the prior domain.

  If the number of transfer candidate feature data 42 belonging to the judgment target node is larger than the classification continuation reference value, the pre-domain determination unit 14 compares the covariance of the judgment target node with the dispersion reference value set in advance. When the covariance of the node to be judged is equal to or less than the dispersion reference value, the prior domain determining unit 14 determines the node to be judged as the prior domain. On the other hand, when the variance of the node to be determined is larger than the variance reference value, the prior domain determination unit 14 determines to further classify the transfer candidate feature data 42 belonging to the node to be determined.

[1.3. Configuration of Prior Domain Evaluation Device 20]
FIG. 3 is a functional block diagram showing the configuration of the preliminary domain evaluation device 20 shown in FIG. As shown in FIG. 3, the prior domain evaluation device 20 includes a temporary storage unit 21, a feature extraction unit 22, a trial transfer learning unit 23, a comparison learning unit 24, and an evaluation unit 25.

  The prior domain evaluation device 20 inputs the target domain 50A stored in the storage device 400, and inputs the prior domain 45 generated by the clustering device 10.

  The temporary storage unit 21 temporarily stores the prior domain 45 input from the clustering device 10.

  The feature extraction unit 22 extracts feature amounts from each of the learning data 51, 51, ... included in the target domain 50A input to the prior domain evaluation device 20, and a plurality of feature data corresponding to each learning data 51 are extracted. Feature data 52 for learning. The learning feature data 52 generated by the feature extraction unit 22 constitutes a target domain 50B.

  The trial transfer learning unit 23 acquires the target domain 50B from the feature extraction unit 22. The trial transfer learning unit 23 acquires, from the temporary storage unit 21, any one of the pre-domains (target pre-domains) of the pre-domains 45 as an evaluation target. The trial transfer learning unit 23 executes machine learning (trial transfer learning) for evaluating the effectiveness of transfer learning of the focused pre-domain, using the acquired target domain 50A and the focused pre-domain. A random forest introduced with transfer learning is used as an algorithm for trial transfer learning. As a result of trial transfer learning, a trial transfer identification unit 63 corresponding to the target pre-domain is generated. The entity of the trial transition identification unit 63 is a data group configured by a plurality of decision trees. The trial transfer identification unit 63 is generated for each prior domain 45.

  The comparison learning unit 24 executes machine learning for comparison (comparison learning) using only the focused pre-domain. A random forest in which transfer learning is not introduced is used as an algorithm for comparison learning. As a result of the comparison learning, the comparison identification unit 64 corresponding to the focused pre-domain is generated. The entity of the comparison identification unit 64 is a data group constituting a plurality of decision trees. The comparison identification unit 64 is generated for each prior domain 45.

  The evaluation unit 25 uses the identification results of the trial transition identification unit 63 and the comparison identification unit 64 to determine whether or not the focused pre-domain is effective for transition learning. The evaluation unit 25 includes a competition value calculation unit 251, a reliability calculation unit 252, and a transition evaluation unit 253.

  The competition value calculation unit 251 compares the identification result of the sample data by the comparison identification unit 64 with the identification result of the sample data by the trial transition identification unit 63. The sample data is any one of learning feature data 52 included in the target domain 50B and transfer candidate feature data 42 included in the target pre-domain. The competition value calculation unit 251 calculates the competition value 251A based on the comparison result. The competition value 251A indicates the degree to which the identification result by the comparison identification unit 64 and the identification result by the trial transfer identification unit 63 do not match.

  The reliability calculation unit 252 calculates the reliability 252A using the identification result of the sample data by the trial transition identification unit 63. The reliability 252A indicates the reliability of the identification result by the trial transition identification unit 63.

  The transfer evaluation unit 253 evaluates whether or not the focused pre-domain is effective for transfer learning based on the competition value 251A and the reliability 252A. The transfer evaluation unit 253 outputs evaluation result data 253A indicating the evaluation of each of the pre-domains 45 to the selective learning device 30.

[1.4. Configuration of Selective Learning Device 30]
FIG. 4 is a functional block diagram showing the configuration of the selection learning device 30 shown in FIG. As shown in FIG. 4, the selection learning device 30 includes a pre-domain selection unit 31, a feature extraction unit 32, and a transfer learning unit 33.

  The prior domain selection unit 31 receives the prior domain 45 from the clustering device 10, and receives the evaluation result data 253A from the prior domain evaluation device 20. The prior domain selection unit 31 selects, from among the prior domains 45 generated by the clustering device 10, the prior domain 45 that has been evaluated to be effective for transfer learning, based on the input evaluation result data 253A.

  The feature extraction unit 32 acquires the target domain 50A stored in the storage device 500. Like the feature extraction unit 22, the feature extraction unit 32 extracts feature amounts from each of the learning data 51, 51,... Included in the acquired target domain 50A, and generates a target domain 50B.

  The transfer learning unit 33 executes machine learning in which transfer learning is introduced, using the target domain 50B and the pre-domain 45 selected by the pre-domain selecting unit 31. The learning algorithm used by the transfer learning unit 33 is the same as the learning algorithm used by the trial transfer learning unit 23. The transfer learning unit 33 generates transfer identification data 80 as a result of machine learning in which transfer learning is introduced.

[2. Outline of operation]
FIG. 5 is a flowchart showing an outline of the operation of the machine learning apparatus 100. As shown in FIG. 5, in the machine learning device 100, the clustering device 10 executes pre-domain generation processing for generating the pre-domain 45 from the transfer candidate data 41, 41,... Stored in the storage device 400 ( Step S1).

  The number of pre-domains 45 generated by the clustering device 10 is not particularly limited. Each of the prior domains 45 has transfer candidate feature data 42 generated by extracting feature amounts from transfer candidate data 41.

  The preliminary domain evaluation device 20 executes a preliminary domain evaluation process to determine whether each of the preliminary domains 45 generated by the clustering device 10 is effective for transfer learning (step S2). The prior domain evaluation device 20 generates evaluation result data 253A as a result of step S2. The evaluation result data 253A is data in which, among the pre-domains 45 generated by the clustering device 10, the pre-domain 45 determined to be effective for transfer learning is specified.

  In the selective learning device 30, the prior domain selection unit 31 selects the prior domain 45 judged to be effective for transfer learning from the prior domains 45 generated by the clustering device 10 based on the evaluation result data 253A ( Step S3).

  The feature extraction unit 32 (see FIG. 4) acquires the target domain 50A from the storage device 500. The feature extraction unit 32 extracts feature amounts from each of the learning data 51 included in the acquired target domain 50A, and generates a plurality of learning feature data 52 (step S4). The process performed by the feature extraction unit 32 is the same as the process performed by the feature extraction unit 22 illustrated in FIG. 3. That is, the feature extraction unit 32 generates a target domain 50B configured of a plurality of learning feature data 52.

  The transfer learning unit 33 executes machine learning in which transfer learning is introduced using the prior domain 45 selected by the prior domain selection unit 31 and the target domain 50B generated by the feature extraction unit 32 (step S5). . The transfer learning unit 33 uses the same learning algorithm (random forest to which transfer learning is introduced) as the learning algorithm used by the trial transfer learning unit 23. This generates transition identification data 80 which is a data group indicating a plurality of decision trees.

  Hereinafter, the reason why the prior domain generation process (step S1) and the prior domain evaluation process (step S2) are performed will be described.

  FIG. 6 is a view showing an example of the distribution of the target domain 50B and the transfer candidate feature data 42. As shown in FIG. FIG. 6 illustrates the distribution of the transfer candidate feature data 42 and the feature data for learning constituting the target domain 50B, taking the case where the number of dimensions of the feature amounts of the transfer candidate feature data 42 and the feature data for learning 52 is two. 52 shows the distribution.

  The target domain 50 B includes learning feature data 52 generated by extracting feature quantities from the learning data 51. Since the plurality of learning data 51 are images including a person photographed at a depression angle of 0 ° as described above, they have characteristics similar to each other. Therefore, in the two-dimensional space shown in FIG. 6, the variation of the learning feature data 52 is small, and the target domain 50B is limited to a relatively narrow region.

  On the other hand, the distribution of the transfer candidate feature data 42 has a larger variation than the learning feature data 52. Since the transfer candidate data 41 is collected by searching for a detection target (person) on the Internet, the imaging conditions of the person of the transfer candidate data 41 are various. The transfer candidate feature data 42 is generated by extracting feature amounts from the transfer candidate data 41. Therefore, the transition candidate feature data 42 is spread over the entire two-dimensional space shown in FIG. 6, and its position is random.

  Here, machine learning in which transfer learning is introduced will be described using an example of detecting a person from an image. In machine learning in which transfer learning is introduced, a target domain and a prior domain are prepared in advance. The target domain is a group of images having features to be detected under predetermined conditions. In the present embodiment, the detection target is a person, and the predetermined condition is that the detection target (person) is included in the image captured at a depression angle of 0 °.

  The pre-domain is a group of images having features to be detected under conditions different from the predetermined conditions described above. The pre-domain is generated by classifying the collected images according to a predetermined rule. For example, when the imaging condition of each acquired image is known, the acquired image can be classified according to the imaging condition. Thus, the pre-domains are a set of images having features common or similar to one another.

  When the machine learning apparatus executes machine learning in which transfer learning is introduced, learning of a pre-domain is performed first, and then learning of a target domain is performed. Then, the machine learning apparatus identifies an image having a feature common or similar to the feature of the person photographed at a depression angle of 0 °, and transfers the feature included in the identified image to the learning result of the image included in the target domain 50B. Let As a result, the number of images constituting the target domain can be reduced, and the identification accuracy of the person can be improved.

  However, if the image features in one prior domain differ significantly from the image features in the target domain, a negative transition occurs. The reason is that the feature of the image in this prior domain is reflected in the learning result of the image in the target domain by transfer learning. As a result, the accuracy of the classifier generated as a result of machine learning that introduced transfer learning is reduced.

  As shown in FIG. 6, when all transition candidate feature data 42 extending over the entire two-dimensional space is one prior domain, the transition candidate feature data 42 apart from the target domain 50B is used for transition learning. become. In this case, the possibility of negative metastasis is very high. In order to prevent the occurrence of negative metastasis, the prior domain 45 is generated by putting together transition candidate feature data 42 having common or similar features, and the generated prior domain 45 introduces the transition learning. It may be determined whether it is effective for machine learning. The pre-domain generation process (step S1) is performed to generate a pre-domain 45 which is a set of transfer candidate feature data 42 having common or similar features.

  FIG. 7 is a view showing an example in which the transfer candidate feature data 42 shown in FIG. 6 is classified. The clustering device 10 generates prior domains 45A to 45G by classifying the transfer candidate feature data 42 illustrated in FIG. 7.

  Among the pre-domains 45A-45G, pre-domains 45A and 45F do not overlap with the target domain 50B. Thus, the pre-domains 45A and 45F are not effective for machine learning that introduced transfer learning. Also, although the pre-domain 45D overlaps with the target domain 50B, the overlapping range is smaller compared to other pre-domains. Thus, the pre-domain 45D can cause negative metastasis and is not effective for metastasis learning.

  Thus, the pre-domain generation process (step S1) may generate a pre-domain (which is not effective for transfer learning) which may cause negative transfer. In order to improve the accuracy of a transition classifier generated as a result of machine learning in which transition learning is introduced, it is desirable to exclude beforehand a pre-domain that is not effective for transition learning. Therefore, the preliminary domain evaluation process (step S2) is performed to identify the preliminary domain effective for the transfer learning among the preliminary domains 45A to 45G generated by the preliminary domain generation process (step S1).

[3. Prior domain generation processing (step S1)]
FIG. 8 is a flowchart of the prior domain generation process (step S1). The operation of the clustering device 10 for generating the pre-domain 45 from the transition candidate data 41, 41,... Stored in the storage device 400 will be described in detail with reference to FIG.

[3.1. Extraction of HOG feature value]
The clustering device 10 acquires all the transfer candidate data 41 stored in the storage device 400. In the clustering device 10, the feature extraction unit 11 (see FIG. 2) extracts the HOG feature amount from each of all the acquired transfer candidate data 41 (step S101). Thereby, a plurality of transfer candidate feature data 42 corresponding to each of all transfer candidate data 41 is generated.

  The feature extraction unit 11 sets, for example, the conditions for extracting the HOG feature amount from the transfer candidate data 41 as follows. The color channel of the transition candidate data 41 is set to grayscale. The size of the transition candidate data 41 is set to 60 pixels vertically and 30 pixels horizontally.

  The number of cells, blocks, and gradient directions are set as parameters at the time of extraction of HOG feature quantities. A cell is a unit area for calculating the gradient direction of luminance. A block is a unit area that creates a histogram in the gradient direction of luminance. The number of gradient directions is the number of divisions in the range of 0 ° to 180 °.

  For example, the size of one cell is set to 5 pixels vertically and 5 pixels horizontally. The size of one block is set to 3 pixels vertically and 3 pixels horizontally. The number of gradient directions is set to nine. When the number of gradient directions is nine, the gradient direction of each cell is divided into nine directions at intervals of 20 °, and is set to one of nine directions. In this case, the number of dimensions of the transfer candidate feature data 42 is 3240.

[3.2. Determination of classification availability at the root node 35R]
FIG. 9 is a diagram showing an initial structure of the classification tree 35 generated by the classification unit 12. The classification unit 12 uses a density forest as an algorithm for classifying the transfer candidate feature data 42. When a density forest is used, a plurality of classification trees are usually generated, but the classification unit 12 generates only one classification tree.

  The classification tree 35 is formed in the process in which the transfer candidate feature data 42 is classified by the classification unit 12. Among the nodes constituting the classification tree 35, nodes satisfying a predetermined condition are determined as the pre-domain.

  The classification unit 12 creates the root node 35R of the classification tree 35 (step S102). The nodes 35A and 35B shown in FIG. 9 are not generated at the time when step S102 is executed. The classification unit 12 inputs all the transfer candidate feature data 42 generated by the feature extraction unit 11 to the root node 35R (step S103). The number of transfer candidate feature data 42 input to the root node 35R is 30,000.

  Next, the advance domain determination unit 14 determines whether all nodes have been selected as classification target nodes in the classification tree 35 (step S104). Since the root node 35R is not selected as a classification target (No in step S104), the prior domain determination unit 14 selects the root node 35R as a classification target (step S105).

  The prior domain determination unit 14 executes step S106 to determine whether or not the root node 35R satisfies the condition as the prior domain. Specifically, the advance domain determination unit 14 acquires the number of transfer candidate feature data 42 belonging to the root node 35R. The pre-domain determining unit 14 determines whether the number of acquired transfer candidate feature data 42 is larger than a preset classification continuation reference value (step S106). The classification continuation reference value is set to 9270, for example.

  The number (30000) of transition candidate feature data 42 belonging to the root node 35R is larger than the classification continuation reference value (9270) (Yes in step S106). In this case, since the number of transfer candidate feature data 42 belonging to the root node 35R is too large, the root node 35R can not be used as the prior domain 45.

  As described above, when one prior domain includes all the transition candidate feature data 42, the accuracy of the transition identification data 80 generated by machine learning in which transition learning is introduced is reduced. Since the root node 35R includes transfer candidate feature data 42 more than the classification continuation reference value, it includes many transfer candidate feature data 42 largely separated from the area of the target domain 50B, as in the case of the one prior domain. In this case, the prior domain determination unit 14 determines that one of the conditions for classifying the transfer candidate feature data 42 belonging to the root node 35R is satisfied.

  The classification continuation reference value is larger than the number of dimensions of the feature quantity extracted by the feature extraction unit 11. For example, in the present embodiment, the classification continuation reference value is set to 9720, which is three times the number of dimensions (3240) of the transition candidate feature data.

  Next, the clustering apparatus 10 executes steps S107 and S108 to determine whether the condition for classifying the transition candidate feature data 42 belonging to the root node 35R is satisfied based on the covariance of the root node 35R. Do.

  The prior domain determination unit 14 instructs the classification unit 12 to calculate the covariance 44 (see FIG. 3) of the node (root node 35R) to be classified. The classification unit 12 outputs the transition candidate feature data 42 belonging to the classification target node (the root node 35R) to the distribution calculation unit 13 according to the instruction of the prior domain determination unit 14. The variance calculation unit 13 uses the transition candidate feature data 42 output from the classification unit 12 to calculate the covariance 13A of the feature amounts of the transition candidate feature data 42 belonging to the node to be classified. The variance calculating unit 13 outputs the calculated covariance 13A to the pre-domain determining unit 14.

  The prior domain determination unit 14 determines whether the covariance 13A (covariance of the root node 35R) calculated by the dispersion calculation unit 13 is larger than a predetermined dispersion reference value (step S108). It is assumed that the covariance 13A is larger than the dispersion reference value (Yes in step S108).

  As described above, the root node 35R includes all the transition candidate feature data 42, and the variation of all the transition candidate feature data 42 is very large. In this case, since the covariance 44 is very large, the prior domain determination unit 14 determines that the transfer candidate feature data 42 belonging to the root node 35R can be further classified. The prior domain determination unit 14 instructs the classification unit 12 to classify the transfer candidate feature data 42 belonging to the root node 35R.

[3.3. Classification of transfer candidate feature data 42]
The classification unit 12 generates nodes 35A and 35B as child nodes of the root node 35R in order to classify the transition candidate feature data 42 belonging to the root node 35R according to an instruction of the prior domain determination unit 14 (step S109). ).

  The classification unit 12 classifies the transition candidate feature data 42 belonging to the root node 35R into one of the nodes 35A and 35B generated in step S109 (step S110). Specifically, the classification destination node of the transfer candidate feature data 42 is determined based on the objective function I shown in the following equation (1).

In Formula (1), S is a parent node (root node 35R). S ^L is a left node of the two child nodes (node 35A), ^{S R} is the right node of the two child nodes (node 35B). Λ (S) is the covariance of the parent node, Λ (S ^L ) is the covariance of the left child node, and Λ (S ^R ) is the covariance of the right child node.

  The classification unit 12 provisionally classifies the transition candidate feature data 42 belonging to the root node 35R in order to calculate the objective function I shown in Equation (1). Specifically, the classification unit 12 sets a provisional branch condition of the transition candidate feature data 42 as follows.

  The number of dimensions of the transfer candidate feature data 42 is 3240. That is, the transfer candidate feature data 42 has 3240 feature quantities. The classification unit 12 randomly selects the k-th (0 ≦ k ≦ 3239) feature amount from the 3240 feature amounts, and randomly sets the threshold value of the k-th feature amount. Thus, a provisional branch condition is set.

  The classification unit 12 provisionally classifies the transition candidate feature data 42 belonging to the root node 35R into the node 35A or 35B based on the set branching condition. The variance calculation unit 13 calculates the covariance of the transition candidate feature data 42 classified into the node 35A and the covariance of the transition candidate feature data 42 provisionally classified into the node 35B. The covariance of the root node 35R has already been calculated in step S105. The classification unit 12 calculates an objective function I of the root node 35R using these three covariances.

  The classification unit 12 sets a plurality of branch conditions in the root node 35R. The classification unit 12 provisionally classifies the transition candidate feature data 42 based on each branch condition in order to calculate an objective function I corresponding to each branch condition. An objective function I in each branch condition is calculated based on the provisionally classified transition candidate feature data 42. The classification unit 12 specifies the largest objective function I among the calculated plurality of objective functions I. The classification unit 12 determines to classify the transition candidate feature data 42 belonging to the root node 35R under the branch condition corresponding to the largest objective function I. Thereby, the transition candidate feature data 42 belonging to the root node 35R is classified into one of the nodes 35A and 35B.

  FIG. 10 is a diagram showing the classification tree 35 after the transfer candidate feature data 42 belonging to the root node 35R is classified. At the time when classification of the transition candidate feature data 42 into the nodes 35A and 35B is finished, child nodes (nodes 35C and 35D) of the node 35B are not generated.

  As a result of classifying 30,000 transition candidate feature data 42 belonging to the root node 35R, 7000 transition candidate feature data 42 are classified into the node 35A. The 23,000 transfer candidate feature data 42 are classified into the node 35B. Thus, step S110 of classifying transition candidate feature data 42 belonging to the root node 35R into two child nodes ends.

3.4. Decision at node 35A]
After the classification of the transition candidate feature data 42 belonging to the root node 35R is completed, the prior domain determination unit 14 determines whether all the nodes have been selected as the classification target (step S104). Since there are nodes 35A and 35B which have not been selected (No in step S104), the pre-domain determining unit 14 selects nodes to be the next determination target in the order of precedence (step S105). Specifically, the classification unit 12 selects the node 35A.

  As shown in FIG. 10, the number of transfer candidate feature data 42 belonging to the node 35A is 7000. Since the number of transfer candidate feature data 42 belonging to the node 35A is equal to or less than the classification continuation reference value (9270) (No in step S106), the prior domain determination unit 14 determines the node 35A as the prior domain 45 (step S111). ). That is, the advance domain determination unit 14 determines not to further classify the transition candidate feature data 42 belonging to the node 35A, and sets the node 35A as a leaf node.

[3.5. Decision at node 35 B]
Next, the prior domain determination unit 14 selects the node 35B as a determination target (step S105). The number of transfer candidate feature data 42 belonging to the node 35B is 23000, which is larger than the classification continuation reference value (9270) (Yes in step S106). Further, it is assumed that the covariance of the node 35B is larger than the dispersion reference value (Yes in step S108). In this case, the prior domain determination unit 14 determines to further classify the transfer candidate feature data 42 belonging to the node 35B.

  The classification unit 12 generates child nodes (nodes 35C and 35D) of the node 35B according to the determination of the prior domain determination unit 14 with respect to the node 35B (step S109). The classification unit 12 classifies the transition candidate feature data 42 belonging to the node 35B into any one of the nodes 35C and 35D, similarly to the classification of the transition candidate feature data 42 in the root node 35R (step S110).

  FIG. 11 is a diagram showing the classification tree 35 after the prior domain generation processing (step S1) is finished. As shown in FIG. 11, as a result of classifying transition candidate feature data 42 belonging to node 35B into nodes 35C and 35D, 15,000 transition candidate feature data 42 are classified into node 35C, and 8,000 transition candidate feature data 42 are classified as node 35D.

  The number of transfer candidate feature data 42 belonging to the node 35C is larger than the classification continuation reference value (9270) (Yes in step S106). Further, it is assumed that the covariance of the node 35C is larger than the dispersion reference value (Yes in step S108). In this case, the prior domain determination unit 14 determines to further classify the transfer candidate feature data 42 belonging to the node 35C. The classification of the transfer candidate feature data 42 belonging to the node 35C will be described later.

  On the other hand, since the number of transfer candidate feature data 42 belonging to the node 35D is equal to or less than the classification continuation reference value (No in step S106), the prior domain determination unit 14 determines the node 35D to be a prior domain.

[3.6. End of classification of transfer candidate feature data 42]
The classification unit 12 generates nodes 35E and 35F as child nodes of the node 35C (step S109), and classifies the transition candidate feature data 42 belonging to the node 35C into the nodes 35E and 35F (step S110).

  The number of transfer candidate feature data 42 belonging to the node 35E is 500, which is equal to or less than the classification continuation reference value (No in step S106). Therefore, the prior domain determination unit 14 determines the node 35E as the prior domain (step S111).

  The number of transfer candidate feature data 42 belonging to the node 35F is 14500, which is larger than the classification continuation reference value (Yes in step S106). On the other hand, it is assumed that the covariance of the node 35F is smaller than the dispersion reference value (No in step S108). In this case, the prior domain determination unit 14 determines that the distribution of the feature amounts of the transfer candidate feature data 42 belonging to the node 35F is very small. For example, it can be considered that most of the transition candidate feature data 42 belonging to the node 35F is generated from the same image. In this case, the prior domain determination unit 14 determines that the transfer candidate feature data 42 included in the node 35F can not be further classified, and determines the node 35F as the prior domain (step S111). As a result, since all nodes constituting the classification tree 35 have been selected as judgment targets (Yes in step S104), the clustering device 10 proceeds to step S112.

[3.7. Advance Domain Exclusion]
The prior domain determination unit 14 confirms the number of transfer candidate feature data 42 possessed by each node determined in the prior domain. If there is a node having transfer candidate feature data 42 whose number is equal to or less than the preset discard reference value, the pre-domain determining unit 14 excludes this node from the pre-domain (step S112). The discarding reference value is set to, for example, the number of dimensions (3240) of the transition candidate feature data 42. Specifically, the node 35E determined to be the pre-domain is excluded from the pre-domain because the number of transfer candidate feature data 42 is 500.

  As mentioned above, if the number of data used for learning is less than the number of dimensions, the accuracy of the generated identification device may be reduced.

  The classification continuation reference value is larger than the number of dimensions of the feature quantity extracted by the feature extraction unit 11. In machine learning, when the number of data used for learning is smaller than the number of dimensions of data used for learning, the learning result of the feature of the data used for learning is overestimated, and the accuracy of the transition identification data 80 decreases. . Therefore, in the present embodiment, the discarding reference value is set to 3240, which is the number of dimensions of the transition candidate feature data 42. As a result, it is possible to prevent the number of transfer candidate feature data 42 belonging to the prior domain 45 from becoming less than the number of dimensions of transfer candidate feature data 42.

  In addition, when the number of transfer candidate feature data 42 included in a certain prior domain is smaller than the discarding reference value, there is a possibility that the transfer candidate feature data 42 included in this prior domain does not have the feature to be detected. high.

  For example, when collecting an image of a person on the Internet, an image in which an object other than the person is photographed may be erroneously acquired as the transfer candidate data 41. The transition candidate feature data 42 generated from the erroneously collected transition candidate data 41 has features different from the transition candidate feature data 42 having human features, and is not effective for transition learning. In addition, since the search condition is an image obtained by shooting a person, it is assumed that the ratio of an image obtained by shooting an object other than a person in the set of transfer candidate data 41 is very small.

  Therefore, when the number of transfer candidate feature data 42 belonging to a certain node is smaller than the discarding reference value, this node is considered to be configured by transfer candidate feature data 42 generated from transfer candidate data 41 collected erroneously. Be The prior domain determination unit 14 excludes, from the prior domain, nodes having transfer candidate feature data 42 in number equal to or less than the discarding reference value.

  As a result, in the classification tree 35 shown in FIG. 11, nodes 35 A, 35 D and 35 F are determined as the pre-domain 45. The clustering device 10 outputs the determined three prior domains 45 to the prior domain evaluation device 20 and the selection learning device 30.

  As described above, the clustering device 10 extracts a feature from each of the transfer candidate data 41 to generate a plurality of transfer candidate feature data 42, and generates a plurality of transfer candidate feature data 42 in the process of creating the classification tree 35. Are classified into the nodes of the classification tree 35. If the number of transfer candidate feature data 42 belonging to a node is less than or equal to the classification continuation reference value or the covariance of transfer candidate feature data 42 belonging to the node is less than or equal to the distribution reference value Decide on. Thereby, it is possible to generate a pre-domain composed of transfer candidate feature data 42 having similar or common features to each other.

[4. Prior domain evaluation process (step S2)]
FIG. 12 is a flowchart of the prior domain evaluation process (step S2) shown in FIG. When the prior domain evaluation device 20 starts the process shown in step S2, the trial transition identification unit 63 is not generated in the trial transition learning unit 23, and the comparison identification unit 64 is generated in the comparison learning unit 24. Not.

[4.1. Generation of target domain 50B]
The prior domain evaluation device 20 acquires the prior domain 45 generated by the clustering device 10. Specifically, the preliminary domain evaluation device 20 acquires three preliminary domains 45 (nodes 35A, 35D, 35F shown in FIG. 11) generated in the process of creating the classification tree 35 shown in FIG. The prior domain evaluation device 20 stores the acquired prior domain 45 in the temporary storage unit 21 (step S201).

  The nodes 35A, 35D, and 35F are hereinafter described as "pre-domain 35A", "pre-domain 35D", and "pre-domain 35F", respectively.

  The feature extraction unit 22 (see FIG. 3) acquires the target domain 50A stored in the storage device 500. The feature extraction unit 22 generates a plurality of learning feature data 52 corresponding to each of the learning data 51 by extracting feature amounts from each of the learning data 51 included in the acquired target domain 50A (Step S202). As a result, a target domain 50B configured of a plurality of learning feature data 52 is generated. The feature extraction unit 22 outputs the generated target domain 50B to the trial transfer learning unit 23.

  The feature extraction unit 22 extracts feature amounts under the same conditions as when the feature extraction unit 11 (see FIG. 2) generates the transition candidate feature data 42 from the transition candidate data 41. Therefore, the number of dimensions of the learning feature data 52 is 3240, which is the same as the number of dimensions of the transfer candidate feature data 42. The reason will be described later.

  The preliminary domain evaluation device 20 selects one of the preliminary domains 45 stored in the temporary storage unit 21 as a target to be evaluated for whether it is effective for transfer learning (step S203). Specifically, the pre-domain 35A is selected first among the pre-domains 35A, 35D, and 35F stored in the temporary storage unit 21.

[4.2. Comparative Learning and Trial Transfer Learning]
The comparison learning unit 24 inputs the prior domain 35A selected in step S203. The comparison learning unit 24 learns the input pre-domain 35A (step S204). The learning algorithm of the comparison learning unit 24 is a random forest in which transfer learning is not introduced. The comparison learning unit 24 generates the comparison identification unit 64 reflecting the learning result of the pre-domain 35A by executing step S204. The comparison identification unit 64 is a data group indicating the structure of a plurality of decision trees.

  The trial transfer learning unit 23 acquires the target domain 50B from the feature extraction unit 22, and acquires the prior domain 35A from the temporary storage unit 21. The trial transfer learning unit 23 performs machine learning in which transfer learning is introduced using the input target domain 50B and the pre-domain 35A (step S205). The learning algorithm of the trial transfer learning unit 23 is a random forest in which transfer learning is introduced. The trial transfer learning unit 23 generates the trial transfer identification unit 63 reflecting the learning result of the target domain 50A and the pre-domain 35A by executing step S205. The trial transition identification unit 63 is a data group indicating the configuration of a plurality of decision trees. Since the learning algorithm and domain used in the trial transfer learning unit 23 are different from those of the comparison learning unit 24, the structure of the trial transfer identifying unit 63 is different from the structure of the comparison identifying unit 64.

[4.3. Evaluation of pre-domain (step S206)]
The evaluation unit 25 uses the trial transfer identification unit 63 generated by the trial transfer learning unit 23 and the comparison identification unit 64 generated by the comparison learning unit 24 so that the prior domain 35A to be evaluated is effective for transfer learning. It is determined whether or not (step S206).

  The evaluation unit 25 calculates two types of parameters, the competition value 251A and the reliability 252A, in order to determine the effectiveness of the transition learning. The evaluation unit 25 uses the identification result of the trial transition identification unit 63 of the data included in the sample group when calculating the reliability 252A. Here, the sample group is a set obtained by combining the learning feature data 52 included in the target domain 50B and the transfer candidate feature data 42 included in the prior domain 35A to be evaluated. Hereinafter, data included in the sample group will be referred to as "sample data". When calculating the competition value 251A, the evaluation unit 25 uses the identification result by the comparison identification unit 64 in addition to the identification result by the trial transition identification unit 63.

[4.3.1. Calculation of the conflict value 251A]
The competition value calculation unit 251 calculates the competition value 251A based on the comparison result of the label of each image generated by the trial transition identification unit 63 and the label of each image generated by the comparison identification unit 64.

  The trial transition identification unit 63 inputs any one of the sample data included in the sample group. The trial transition identification unit 63 performs identification processing of a person on the sample data, and generates a label 73 indicating the identification result. The value of the label 73 is, for example, 0 or 1. When the label 73 is 0, the label 73 indicates that the sample data does not include human features. When the label 73 is 1, the label 73 indicates that the sample data includes the feature of a person. The trial transition identification unit 63 outputs the generated label 73 to the conflict value calculation unit 251.

  The trial transition identifying unit 63 calculates not only the label 73 but also the probability 83 indicating the likelihood of the label 73 as the identification result of the sample data. The accuracy 83 is used to calculate the reliability 252A described later.

  The comparison identification unit 64 inputs the same data as the sample data input to the trial transition identification unit 63. The comparison identification unit 64 performs identification processing of a person on the sample data, and generates a label 74 indicating the identification result. The value of the label 74 is 0 or 1 similarly to the label 73. If the label 74 is 0, the label 74 indicates that the sample data does not contain human features. When the label 74 is 1, the label 74 indicates that the sample data contains the feature of a person. The comparison identification unit 64 outputs the generated label 74 to the conflict value calculation unit 251.

  The competition value calculation unit 251 calculates the competition value 251A using the labels 73 and 74 generated from the sample data. The competition value 251A is calculated by the following equation (2).

In the equation (2), E _c1 represents the contention value 251A. X indicates a sample group. x indicates an element (sample data) constituting a sample group. M (x) indicates the label 74 generated from the element x. T (x) indicates the label 73 generated from the element x. [M (x) ≠ T (x)] indicates the number of sample data for which the label 74 and the label 73 did not match. | X | is the number of elements constituting the sample group X.

  The conflict value 251A calculated by the equation (2) indicates the probability that the labels 73 and 74 generated from the same sample data do not match. The competition value 251A is a numerical value of 0 or more and 1 or less. As the competition value 251A approaches 0, the competition value 251A indicates that the effectiveness of the pre-domain 35A in transfer learning is high. On the other hand, the closer the competition value 251A is to 1 the more. It shows that the effectiveness of the pre-domain 35A in transfer learning is low.

  If there are many differences between the learning feature data 52 included in the target domain 50B and the transfer candidate feature data 42 included in the pre-domain 35A to be evaluated, the pre-domain 35A is not effective for transfer learning. In this case, the contention value 251A approaches one. The reason will be described below.

  As described above, the comparison learning unit 24 learns only the pre-domain 35A. Therefore, only the learning result of the prior domain 35A is reflected in the comparison identification unit 64.

  On the other hand, the trial transfer identification unit 63 executes machine learning in which transfer learning is introduced using the target domain 50A and the pre-domain 35A. However, when there are many differences between the learning feature data 52 included in the target domain 50B and the transfer candidate feature data 42 included in the evaluation target prior domain 35A, the learning of transfer candidate feature data 42 included in the advance domain 35A The result is not reflected in the learning result of the learning feature data 52. That is, it can be considered that the trial transfer identification unit 63 and the comparison identification unit 64 are generated by learning different domains. In this case, the case where the identification results of the trial transition identification unit 63 and the comparison identification unit 64 do not match increases, and the competition value 251A increases. Therefore, based on the competition value 251A, it is possible to determine whether the prior domain 35A is effective for transfer learning.

[4.3.2. Calculation of reliability]
The reliability calculation unit 252 calculates the reliability 252A based on the label 73 and the accuracy 83 of each image generated by the trial transition identification unit 63. In the calculation of the reliability 252A, the identification result of the sample data by the comparison identification unit 64 is not used.

  As described above, the trial transition identification unit 63 generates the label 73 indicating the identification result of the person with respect to the sample data, and the accuracy 83 indicating the likelihood of the label 73. The certainty 83 is a value of 0 or more and 1 or less, and as the certainty 83 approaches 1, the possibility that the label 73 is erroneous decreases.

  The reliability calculation unit 252 inputs the label 73 and the accuracy 83 of each sample data from the trial transition identification unit 63. The reliability calculation unit 252 calculates the reliability 252A by calculating the following equation (3) using the label 73 and the accuracy 83 of each input sample data.

In the above equation (3), E _c2 represents the reliability 252A. x shows the element (sample data) which comprises the sample group X similarly to said formula (2). | X | is the number of elements of the sample group X. P _T (x) indicates the accuracy 83 of the element x. P _T (x) is the average of the probability of the class set to the leaf node that the sample data has reached in each decision tree when the sample data is input to each decision tree constituting the trial transition identification unit 63 . T (x) indicates the label 73 of the element x. y is a label (y = 1) indicating the presence of a person. That is, the reliability 252A is a value obtained by dividing the total value of the certainty 83 calculated when the label 73 matches the label y by the number of elements of the sample group X. The reliability 252A is a value of 0 or more and 1 or less, and the closer to 1, the higher the effectiveness of the prior domain 35A in transfer learning.

  When the transfer candidate feature data 42 of the prior domain 35A has a feature amount similar to the feature amount of the learning feature data 52, the trial transfer learning unit 23 learns the transfer candidate feature data 42 by trial transfer learning. The result is transferred to the learning result of the learning feature data 52. The trial transition identification unit 63 reflects the learning results of the learning feature data 52 and the transition candidate feature data 42 of the pre-domain 35A. When the trial transition identification unit 63 performs identification processing on each data of the sample group used for trial transition learning, the label 73 is considered to be 1 and the certainty 83 also approaches 1. Therefore, the reliability 252A approaches 1 when the learning feature data 52 and the transition candidate feature data 42 of the pre-domain 35A are similar (when the pre-domain 35A is effective in transfer learning).

[4.3.3. Evaluation of prior domain by metastasis evaluation unit 253}
The transition evaluation unit 253 inputs the competition value 251A and the reliability 252A. The transfer evaluation unit 253 evaluates the effectiveness of the pre-domain 35A in transfer learning based on the entered competition value 251A and the reliability 252A.

  The transfer evaluation unit 253 calculates the comprehensive evaluation value using the following equation (4).

  In the equation (4), E is a comprehensive evaluation value obtained from the competition value 251A and the reliability 252A. As the effectiveness in transfer learning of the pre-domain 35A decreases, the competition value 251A increases. On the other hand, the reliability 252A decreases in reverse. In order to adjust the tendency of the reliability 252A to the tendency of the competition value 251A, a value obtained by subtracting the reliability 252A from 1 is used for the calculation of the comprehensive evaluation value.

  The comprehensive evaluation value calculated by the above equation (4) is a value of 0 or more and 1 or less, and approaches 0 as the effectiveness of transfer learning becomes higher. The transfer evaluation unit 253 determines that the pre-domain 35A is effective in transfer learning, when the calculated comprehensive evaluation value is smaller than a preset threshold value.

[4.4. Specify the following pre-domains]
After the evaluation of the effectiveness of the pre-domain 35A in the transition learning (step S206) is completed, the trial transition identifying unit 63 and the comparison identifying unit 64 used for the evaluation of the pre-domain 35A are deleted (step S207) . This is because the trial transition identification unit 63 and the comparison identification unit 64 corresponding to the pre-domain 35A are not used in the evaluation of the effectiveness of other pre-domains in transfer learning.

  The preliminary domain evaluation device 20 determines whether all the preliminary domains stored in the temporary storage unit 21 have been selected (step S208). When all the preliminary domains have not been selected (No in step S208), the preliminary domain evaluation device 20 returns to step S203 to evaluate the effectiveness in transfer learning of the non-selected preliminary domains. This evaluates the effectiveness of the pre-domains 35D and 35F in transfer learning.

[4.5. Generation of evaluation result data 253A]
When all pre-domains have been selected (Yes in step S208), the transfer evaluation unit 253 creates evaluation result data 253A indicating the evaluation results of each of the pre-domains 35A, 35D, and 35F. The number of prior domains judged to be effective for transfer learning is not particularly limited. The transfer evaluation unit 253 outputs the generated evaluation result data 253A to the selective learning device 30.

  Refer again to FIG. In the selective learning device 30, the prior domain selection unit 31 selects, from among the prior domains 45 generated by the clustering device 10 based on the evaluation result data 253A, the prior domains 35A, 35D and 35F determined to be effective for transfer learning. Is selected (step S3). The feature extraction unit 32 (see FIG. 4) acquires the target domain 50A from the storage device 500, and extracts a feature amount from each of the learning data 51 included in the acquired target domain 50A (step S4). Thus, a target domain 50B including the learning feature data 52 is generated. The feature extraction unit 32 extracts feature amounts under the same conditions as when the feature extraction unit 22 (see FIG. 2) extracts feature amounts from the learning data 51.

  The transfer learning unit 33 uses the selected prior domains 35A, 35D and 35F and the target domain 50B generated by the feature extraction unit 32 to execute machine learning in which transfer learning is introduced (step S5). This generates transition identification data 80 which is a data group indicating a plurality of decision trees.

  As described above, the machine learning apparatus 100 extracts features from the transfer candidate data 41, 41,... Stored in the storage device 400, and generates transfer candidate feature data 42, 42,. The machine learning apparatus 100 classifies the transition candidate feature data 42, 42, ... into a plurality of groups based on the extracted feature quantities. The machine learning apparatus 100 determines whether to determine the classified group as the pre-domain based on the number or covariance of the transfer candidate feature data 42 in the classified group. Thereby, the prior domain used for transfer learning can be efficiently generated from transfer candidate data 41.

[Modification]
In the above embodiment, when the clustering device 10 classifies the transition candidate feature data 42, the case of generating a binary tree as the classification tree 35 using the density forest has been described as an example, but the present invention is not limited thereto. . The clustering device 10 may classify the transition candidate feature data 42 using another classification algorithm such as the k-means method. In this case, the number of child nodes created in step S109 (see FIG. 8) may be three or more.

  In addition, the clustering device 10 may classify the transition candidate feature data 42 using two or more classification algorithms. For example, the clustering device 10 performs classification algorithm based on whether the number of transfer candidate feature data 42 belonging to the classification target node is larger than a reference value (algorithm change reference value) for determining a change in the classification algorithm. Decide.

  FIG. 13 is a diagram showing an example of the classification tree 35 generated using the k-means method and the density forest. For example, assume that the algorithm change reference value is set to 25000.

  The number of transfer candidate feature data 42 belonging to the root node 35R is 30,000, which is larger than the algorithm change reference value. In this case, the clustering device 10 generates nodes 36A, 36B and 36C as child nodes of the root node 35R. Then, the clustering device 10 generates the transition candidate feature data 42 belonging to the root node 35R using the k-means method to generate nodes 36A, 36B, and 36C.

  Then, the numbers of transfer candidate feature data 42 belonging to the nodes 36A and 36C are 5000 and 8000, respectively, which are below the classification continuation reference value (9270). The clustering device 10 determines nodes 36A and 36C as pre-domains, respectively. On the other hand, the number of transfer candidate feature data 42 belonging to the node 36 B is 17000, which is larger than the classification continuation reference value. In this case, the clustering device 10 further classifies the transition candidate feature data 42 belonging to the node 36B.

  Since the number (17000) of transition candidate feature data 42 belonging to the node 36B is less than or equal to the algorithm change reference value (25000), the clustering apparatus 10 uses density forest for classification of the transition candidate feature data 42 belonging to the node 36B. Decide. The clustering device 10 generates nodes 36D and 36E as child nodes of the node 36B, and classifies transition candidate feature data 42 belonging to the node 36B.

  As described above, by switching the classification algorithm in accordance with the number of transfer candidate feature data 42 belonging to the classification target node, the transfer candidate feature data 42 can be classified at high speed.

  Moreover, although the example which the selective learning apparatus 30 (refer FIG. 4) equips with the feature extraction part 32 was demonstrated in the said embodiment, it is not restricted to this. The selective learning device 30 may generate the transition identification data 80 using the target domain 50B generated by the feature extraction unit 22 included in the prior domain evaluation device 20 (see FIG. 3). Alternatively, the preliminary domain evaluation device 20 may extract feature amounts from the transfer candidate data 41 corresponding to each of the preliminary domains 45 to generate transfer candidate feature data 42. Alternatively, the selective learning device 30 may generate the transfer candidate feature data 42 by extracting feature amounts from the transfer candidate data 41 corresponding to the pre-domain determined to be effective for transfer learning.

  In any case, the transition candidate feature data 42 used in each of the clustering device 10, the prior domain evaluation device 20, and the selection learning device 30 are all generated by extracting feature amounts from the transition candidate data 41 under the same conditions. Is desirable. Similarly, it is desirable that the learning feature data 52 be generated by extracting feature quantities from the learning data 51 under the same conditions. The reason will be described below.

  For example, when the extraction condition of the feature amount differs between the clustering device 10 and the pre-domain evaluation device 20, the transfer candidate feature data 42 generated in the clustering device 10 has a distribution in the transfer candidate feature data 42 in the pre-domain evaluation device 20. Have different distributions. The positional relationship between the target domain and the prior domain is different between the transition candidate feature data 42 generated in the clustering device 10 and the transition candidate feature data 42 in the prior domain evaluation device 20. As a result, in the prior domain evaluation device 20, the accuracy of determining whether the prior domain generated by the clustering device 10 is effective in the transfer learning is lowered.

  The distribution of transfer candidate feature data 42 in the prior domain 45 determined to be effective by the prior domain evaluation device 20 changes similarly even when the extraction conditions of the feature amount differ between the prior domain evaluation device 20 and the selection learning device 30. . As a result, the learning accuracy of machine learning in which transfer learning is introduced in the selective learning device 30 may be reduced, and the identification accuracy of a person using the transfer identification data 80 may be reduced.

  On the other hand, the transition identification data 80 is generated with accuracy when evaluating the effectiveness of the pre-domain by aligning the extraction conditions of the feature amounts in the clustering device 10, the pre-domain evaluation device 20, and the selection learning device 30. The accuracy of learning can be prevented from decreasing.

  Although the trial transfer learning unit 23, the comparison learning unit 24, and the transfer learning unit 33 use a random forest as a learning algorithm in the above embodiment, the present invention is not limited to this. For example, the trial transfer learning unit 23, the comparison learning unit 24, and the transfer learning unit 33 may use various algorithms such as ID3 (Iterative Dichotomizer 3), boosting, and a neural network. Regardless of which learning algorithm is used, the trial transfer learning unit 23 and the transfer learning unit 33 execute machine learning in which transfer learning is introduced, and the comparison learning unit 24 performs machine learning in which transfer learning is not introduced. do it.

  In the above embodiment, although the example in which the transition evaluation unit 253 calculates the comprehensive evaluation value by multiplying the competition value 251A and the reliability 252A has been described, the present invention is not limited thereto. For example, the transfer evaluation unit 253 may calculate the sum of the competition value 251A and the reliability 252A as a comprehensive evaluation value. That is, the transfer evaluation unit 253 may calculate the comprehensive evaluation value using the competition value 251A and the reliability 252A.

  In the above embodiment, although the machine learning device 100 extracts the HOG feature value from each of the transfer candidate data 41 and the learning data 51 has been described as an example, the present invention is not limited to this. For example, when learning the face of a person, the machine learning apparatus 100 may extract a Haar-like feature. The machine learning apparatus 100 may appropriately change the feature quantities extracted from the transfer candidate data 41 and the learning data 51 according to the learning target.

  Although the machine learning device 100 has described the example in which the transition identification data 80 for detecting a person is generated in the above embodiment, the present invention is not limited to this. The target of learning may be measurement data measured by a sensor. The type of sensor is not particularly limited, and various measurement data such as an acceleration sensor and an optical sensor can be used. For example, machine learning may be performed to use the measurement data of these sensors to provide automated driving of the vehicle.

  A part or all of the machine learning apparatus 100 according to the above-described embodiment may be realized as an integrated circuit (for example, an LSI, a system LSI, etc.).

  Part or all of the processing of each functional block in the above embodiment may be realized by a program. And a part or all of processing of each functional block of the above-mentioned embodiment is performed by central processing unit (CPU) in a computer. Further, programs for performing respective processing are stored in a storage device such as a hard disk or a ROM, and are read out from the ROM or to the RAM and executed.

  Further, each process of the above-described embodiment may be realized by hardware or may be realized by software (including a case where it is realized with an OS (Operating System), middleware, or a predetermined library). Furthermore, it may be realized by mixed processing of software and hardware. Needless to say, when the machine learning apparatus 100 according to the above-described embodiment is realized by hardware, it is necessary to perform timing adjustment for performing each process. In the above embodiment, for the convenience of description, the details of the timing adjustment of various signals generated in the actual hardware design are omitted.

  Further, the order of execution of the processing method in the above embodiment is not necessarily limited to the description of the above embodiment, and the order of execution can be interchanged without departing from the scope of the invention.

  A computer program that causes a computer to execute the above-described method and a computer readable recording medium recording the program are included in the scope of the present invention. Here, examples of the computer-readable recording medium include a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD, a DVD-ROM, a DVD-RAM, a large capacity DVD, a next-generation DVD, and a semiconductor memory. .

  The computer program is not limited to one recorded in the recording medium, but may be transmitted via a telecommunication line, a wireless or wired communication line, a network represented by the Internet, or the like.

100 machine learning device 10 clustering device 11, 22, 32 feature extraction unit 12 classification unit 13 distributed calculation unit 14 prior domain determination unit 20 prior domain evaluation device 21 temporary storage unit 23 trial transition learning unit 24 comparison learning unit 25 evaluation unit 251 competition Value calculation unit 252 reliability calculation unit 253 transition evaluation unit 30 selective learning device 31 prior domain selection unit 33 transition learning unit

Claims (3)

A machine learning apparatus that executes machine learning in which transfer learning is introduced to learn a detection target,
A clustering device that classifies a plurality of transfer candidate data used for the machine learning and generates a pre-domain used for the machine learning;
A pre-domain evaluation device that evaluates whether or not the pre-domain generated by the clustering device is effective for the machine learning;
Equipped with
The clustering device
A feature extraction unit for clustering which extracts features from each of the plurality of transfer candidate data and generates a plurality of transfer candidate feature data;
A classification unit that classifies each transfer candidate feature data into a plurality of groups including a first group and a second group based on the features of each of the plurality of transfer candidate feature data generated by the clustering feature extraction unit;
When the number of transfer candidate feature data classified into the first group by the classification unit is equal to or less than a predetermined classification continuation reference value, the first group is determined to be a pre-domain used for the machine learning, and the transfer candidate A pre-domain determination unit that determines to further classify transfer candidate feature data classified into the first group if the number of feature data is larger than the classification continuation reference value;
Equipped with
The prior domain evaluation device
When the first group is determined to be the prior domain by the prior domain determination unit, for learning having transfer candidate feature data included in the first group, and each having a feature to be detected under a predetermined condition A trial transfer learning unit that executes the machine learning using a target domain including data to generate an evaluation discriminator for evaluating the pre-domain;
A determination unit that determines whether the first group is effective for the machine learning based on the trial transfer identification unit generated by the trial transfer learning unit;
Machine learning device comprising: The machine learning device according to claim 1 , wherein
The pre-domain evaluator may further:
A learning feature extraction unit that extracts features of each of the learning data included in the target domain and generates learning feature data;
Equipped with
The trial transfer learning unit executes the machine learning using the learning feature data,
The machine learning apparatus, wherein a condition for the feature extraction unit for learning to extract a feature from data for learning is the same as a condition for the feature extraction unit for clustering to extract a feature from each of the plurality of transfer candidate data. The machine learning apparatus according to claim 2 , further comprising:
A selection learning device that executes the machine learning using the target domain and all pre-domains determined to be effective for the machine learning by the pre-domain evaluating device to generate a transition identification unit;
Machine learning device comprising:

Images