WO2016111241A1 - 学習装置、識別器、学習方法および記録媒体 - Google Patents
学習装置、識別器、学習方法および記録媒体 Download PDFInfo
- Publication number
- WO2016111241A1 WO2016111241A1 PCT/JP2016/000015 JP2016000015W WO2016111241A1 WO 2016111241 A1 WO2016111241 A1 WO 2016111241A1 JP 2016000015 W JP2016000015 W JP 2016000015W WO 2016111241 A1 WO2016111241 A1 WO 2016111241A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- dictionary
- labeled
- loss
- sample
- update
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/242—Dictionaries
Definitions
- the present invention relates to a learning device, a discriminator, a learning method, and a recording medium.
- Non-Patent Document 1 describes a method of stopping active learning based on stability prediction.
- Patent Document 1 describes that a label is predicted for a sample (content) to which no label is attached, and the certainty factor of the predicted label (predicted label) is calculated.
- the certainty factor of the predicted label indicates the certainty of the predicted label.
- Patent Document 1 describes that active learning is terminated when the certainty factor of a predicted label is equal to or greater than a predetermined threshold.
- the above-described technique has a problem that a lot of processing time is required because new learning and evaluation for determining the end of active learning occur.
- the present invention has been made in view of the above problems, and an object thereof is to provide a technique for further improving the man-hours and processing time for learning.
- a learning apparatus includes an updating unit that updates a dictionary used for a classifier, a dictionary updated by the updating unit, one or more samples, and a labeled sample to which a label is attached. And calculating means for calculating a ratio with respect to the number of labeled samples as a loss with respect to the whole labeled sample, and determining means for determining whether or not to update the dictionary using the loss.
- the update means updates the dictionary using the labeled sample to which a new labeled sample is added when the determination means determines that the dictionary is to be updated, and the determination means The loss calculated using the updated dictionary and the labeled sample before adding the new labeled sample using the dictionary before update Using loss and calculated to the body, and it is determined whether to update the dictionary.
- the learning method updates a dictionary used for a discriminator, and uses the updated dictionary and one or more samples and labeled samples to which labels are given, A ratio with respect to the number of labeled samples is calculated as a loss with respect to the entire labeled sample, and using the loss, it is determined whether to update the dictionary, and it is determined to update the dictionary. And updating the dictionary using a labeled sample to which the new labeled sample is added, and using the updated dictionary and the labeled sample to which the new labeled sample is added.
- the ratio to the number of labeled samples is calculated as the loss for the whole labeled sample, the loss calculated using the updated dictionary, and the pre-update Using a calculated loss for the entire labeled sample prior to the addition of the new labeled samples using the dictionary, and determines whether to update the dictionary.
- a classifier that identifies data using a dictionary that is determined not to be updated by the learning device is also included in the scope of the present invention.
- a computer program that realizes the learning device or the learning method by a computer and a computer-readable recording medium in which the computer program is stored are also included in the scope of the present invention.
- the man-hours and processing time for learning can be further improved.
- FIG. 1 is a functional block diagram illustrating an example of a functional configuration of the learning device 100 according to the present embodiment.
- the direction of the arrow in a drawing shows an example and does not limit the direction of the signal between blocks.
- the directions of the arrows in the drawings show an example and do not limit the direction of signals between the blocks.
- the learning device 100 includes a selection unit 101, an acquisition unit 102, an update unit 103, a calculation unit 104, a comparison unit (determination unit) 105, and an output unit 106. And a storage unit 107.
- the number of samples (learning data) used for learning and not labeled is M (M is a natural number).
- the number of samples to which labels are assigned is N (N is a natural number).
- the label indicates a class to which each sample belongs. That is, a sample to which a label is assigned can be said to be a sample (labeled sample) to which a correct class (referred to as correct class) label is assigned.
- a sample to which a correct class label is assigned is called a labeled sample
- a sample to which no label is given is called an unlabeled sample.
- the correct class label is also called a correct label.
- the learning data (sample) input to the learning device 100 is described as being stored in the storage unit 107, but the learning data is directly input to each member described later. It may be done.
- the selection unit 101 selects a sample to be labeled from the learning data (sample) input to the learning device 100 stored in the storage unit 107 in accordance with an instruction from the comparison unit 105 described later. That is, the selection unit 101 selects a sample to be labeled from samples that are used for learning and that are not labeled. Assuming that the set of learning data is a sample set C, a set of unlabeled samples is a sample set B, and a set of labeled samples is a sample set A, the sample set C is the sum of the sample set A and the sample set B. It becomes a set. When a label is assigned to a sample belonging to the sample set B, the sample to which the label is assigned belongs to the sample set A. Therefore, the number of samples belonging to the sample set B is reduced by one.
- the sample selection method performed by the selection unit 101 will be described later.
- the storage unit 107 stores learning data input to the learning device 100.
- the storage unit 107 stores each sample in a state in which information indicating whether each sample belongs to the sample set A or the sample set B is associated.
- the storage unit 107 stores each of one or more samples associated with information indicating that it belongs to the sample set A in a state in which the label given to the sample is known.
- the storage unit 107 stores a sample associated with information on a set.
- the sample whose information related to the set indicates the sample set A is associated with a label.
- the storage unit 107 may store a dictionary (parameter) to be described later, loss of a labeled sample, and the like.
- the storage unit 107 may be built in the learning device 100 or may be realized by a storage device separate from the learning device 100.
- the learning data, the parameters, and the loss are stored in the same storage unit, but they may be stored in different storage units.
- the acquisition unit 102 acquires information regarding the sample to which the label is attached. Specifically, when the user assigns a label to the sample selected by the selection unit 101, the acquisition unit 102 acquires information regarding the sample to which the label is assigned. For example, when the user gives a label to a sample without a label using an input device (not shown), the acquisition unit 102 gives a sample (labeled sample) to which a label is given based on operation information of the input device. You may acquire the information about. The acquisition unit 102 acquires information about the sample in a state where it can be seen what label is given to each sample.
- the sample to which the label is assigned belongs to the sample set B before the label is assigned. Therefore, the acquisition unit 102 deletes the sample with the label from the sample set B and includes the sample set A in the sample set A. That is, the acquisition unit 102 updates the information on the set associated with the sample stored in the storage unit 107 from the information on the sample set B to the information indicating the sample set A. Then, the acquisition unit 102 associates a label with the sample stored in the storage unit 107.
- the acquisition part 102 outputs the information which shows having acquired the information regarding the sample to which the label was given to the update part 103.
- the acquisition unit 102 may output information indicating that the information associated with the sample stored in the storage unit 107 has been updated to the update unit 103.
- the acquisition unit 102 may output the number N of labeled samples to the update unit 103 as information indicating that the information related to the sample with the label is acquired.
- the acquisition unit 102 is configured to acquire a sample with a new label or all samples with a label (labeled samples belonging to the sample set A) from the storage unit 107 and supply them to the update unit 103. There may be.
- the update unit 103 receives, from the acquisition unit 102, information indicating that information about a sample to which a label is attached, a sample to which a new label is attached, or all samples to which a label is attached.
- the update unit 103 acquires a labeled sample from the storage unit 107 when receiving information indicating that information related to a sample to which a label has been added or a sample having a new label has been received.
- the updating unit 103 updates itself and / or the dictionary stored in the storage unit 107 using the labeled sample belonging to the sample set A.
- the dictionary is a parameter used by the classifier to identify predetermined data such as sound and image. Further, the parameter is not particularly limited, but, for example, indicates a characteristic of a typical pattern that is identified as a specific one by a classifier. The discriminator identifies unknown data based on this parameter.
- the dictionary update method will be described later.
- the update unit 103 outputs the updated dictionary (identifier parameter) to the calculation unit 104 and the output unit 106.
- the calculation unit 104 receives parameters from the update unit 103. In addition, the calculation unit 104 acquires a labeled sample from the storage unit 107. Then, the calculation unit 104 calculates the loss of the labeled sample using the parameter and the labeled sample stored in the storage unit 107.
- n 1,.
- the label t n is a label assigned to the labeled sample x n and represents the correct answer class.
- the loss for the entire labeled sample is represented by an equation having the number of labeled samples (N) in the denominator. That is, the loss for the entire labeled sample is obtained by calculating the ratio of the sum of loss (x; ⁇ ) to the number of labeled samples.
- Equation (1) ⁇ is a dictionary (a parameter of a discriminator). Further, loss (x; ⁇ ) represents a loss (error condition) with respect to the vector x when the parameter ⁇ is used.
- loss (x; ⁇ ) is defined as in the following equation (2).
- K is the number of classes
- 1 (•) is an instruction function that returns a predetermined value in accordance with a true / false value.
- this indicator function returns 1 if the conditional expression in parentheses is true, and returns 0 if it is false.
- J is a natural number of 1 or more and K or less.
- r kj (x n ; ⁇ ) is a measure representing the ease of mistakes. If r kj (x n ; ⁇ ) is negative, it indicates positive recognition, and if it is positive, it indicates erroneous recognition.
- loss (x; ⁇ ) is a loss for each labeled sample calculated using the dictionary updated by the updating unit 103. This r kj (x n ; ⁇ ) is defined by the following equation (3).
- g k (x n ; ⁇ ) represents an identification function of class ⁇ k .
- the value of the discrimination function is positive.
- This g k (x n ; ⁇ ) is defined such that the class has a higher value as the class is more likely to belong. Therefore, the class ⁇ k that maximizes g k (x n ; ⁇ ) is the class (identification result) determined by the classifier.
- f (•) is a positive monotonically increasing function that determines the amount of loss with respect to easy mistake, and is defined by the following equation (4), for example.
- ⁇ ( ⁇ > 0) is a parameter representing the inclination. For samples that are misrecognized, a large loss is given.
- the calculation unit 104 calculates the loss L N of the labeled sample using the parameter and the labeled sample stored in the storage unit 107.
- the calculation unit 104 may acquire the number N of labeled samples used for calculating the loss L N by counting the number of labeled samples stored in the storage unit 107, or may update the update unit from the acquisition unit 102.
- the information may be acquired via 103.
- the calculation unit 104 outputs the loss LN of labeled samples associated with the number of labeled samples (in this case, N) when calculating the loss to the comparison unit 105.
- the calculation unit 104 may store the calculated loss LN in the storage unit 107 in association with the number of labeled samples when the loss is calculated.
- the comparison unit 105 receives the loss LN of labeled samples associated with the number of labeled samples from the calculation unit 104.
- the comparison unit 105 sets the received loss L N as the loss L new, and is the loss of the labeled samples having a number smaller than the number of labeled samples associated with the received loss L N.
- the loss calculated by the calculation unit 104 is referred to as a loss L old .
- the loss of the labeled samples that is smaller than the number of labeled samples associated with the received loss L N may be stored in the comparison unit 105 or may be stored in the storage unit 107.
- the comparison unit 105 determines whether or not to update the dictionary by comparing the loss L new and the loss L old . When the loss L old is larger than the loss L new , the comparison unit 105 determines not to update the dictionary and outputs an instruction to output the dictionary to the output unit 106. Further, when the previously calculated loss L old is equal to or less than the received loss L new , the comparison unit 105 determines that the dictionary is to be updated, and gives an instruction to select a sample to be labeled. 101.
- comparing section 105 sets loss L N when the number of labeled samples is N as loss L new and sets loss L N-1 when the number of labeled samples is N ⁇ 1 as loss L. As old , the loss L new and the loss L old are compared.
- the output unit 106 receives an instruction to output a dictionary from the comparison unit 105. Further, the output unit 106 receives the dictionary (parameter) updated by the update unit 103 from the update unit 103. And the output part 106 outputs the dictionary received from the update part 103 based on the instruction
- This dictionary is a parameter used when the calculation unit 104 calculates the loss. Note that the output unit 106 may acquire the dictionary to be output from the storage unit 107 and output it.
- FIG. 2 is a diagram illustrating an example of a change in loss of labeled samples with respect to an increase in the number of labeled samples in the learning apparatus 100 according to the present embodiment.
- the horizontal axis indicates the number of labeled samples
- the vertical axis indicates the loss of labeled samples.
- Expression (5) is an expression obtained by simplifying Expression (1) and omits the notation of ⁇ from Expression (1).
- m is an arbitrary natural number between 1 and N.
- n is m + 1 greater than the value of the loss (x n) it is a loss for x n is, when a sufficiently small value, the second term of the right side of the equation (5) becomes sufficiently small value. Therefore, the first term is dominant on the right side of Equation (5).
- the loss L N decreases in inverse proportion to N.
- the loss LN can be an index for determining the end of the work of labeling the correct answer class.
- the active learning is performed so that a label is preferentially given to a sample with a low reliability of the identification result. But less), the loss spikes.
- the loss starts to decrease in inverse proportion to the increase in the number of labeled samples from the point when the number of labeled samples exceeds a certain value (TH in FIG. 2). .
- FIG. 3 shows that the learning apparatus 100 according to the present embodiment recognizes the evaluation set with a label using the recognition dictionary obtained each time the number of labeled samples increases (called a stage). It is a figure which shows an example of the error rate at the time.
- the horizontal axis indicates the number of labeled samples
- the vertical axis indicates the error rate.
- TH is the same value as the number of labeled samples (TH) indicated by the broken line in FIG. 2, and when the loss starts to decrease in inverse proportion to the number of labeled samples in FIG. The number of labeled samples is shown.
- the error rate sharply decreases at the initial stage (when the number of labeled samples is smaller). Then, after the number of labeled samples is a certain value (TH), the error rate is substantially constant.
- the comparison unit 105 determines whether or not the number N of labeled samples when the calculation unit 104 calculates the loss L N is the value TH shown in FIG. That is, it can be said that the comparison unit 105 determines a time point at which the loss LN starts to decrease in inverse proportion to the number of labeled samples.
- FIG. 4 shows an example of the loss calculated using an expression in which 1 / N on the right side of Expression (1) is omitted in order to explain the effect of calculating the loss using Expression (1).
- FIG. 4 is a diagram illustrating another example of a change in loss of labeled samples with respect to an increase in the number of labeled samples.
- the horizontal axis indicates the number of labeled samples
- the vertical axis indicates the loss of labeled samples.
- TH is the same value as the number of labeled samples (TH) indicated by a broken line in FIG.
- the equation used when the calculation unit 104 of the learning device 100 according to the present embodiment calculates the loss is an equation having the number of labeled samples (N) in the denominator as shown in Equation (1). It can be seen that it is preferable.
- FIG. 5 is a flowchart showing an example of the processing flow of the learning apparatus 100 according to the present embodiment.
- the identifier identifies the input data to the k-th class omega k.
- the discriminant function is not limited to the formula shown in formula (6), and an arbitrary discriminant function may be used. Even in this case, the learning apparatus 100 can preferably execute the processing described below.
- the sample to which the correct class label is assigned is stored in the storage unit 107 as a labeled sample.
- These labeled samples belong to the sample set A described above.
- the loss L N calculated by the calculation unit 104 using the equation (1) is not calculated at this time. Therefore, the value of the loss L old calculated before is 0.
- the loss L old is stored in the comparison unit 105. As described above, the loss L old may be stored in the storage unit 107.
- the storage unit 107 stores, as learning data, samples to which M labels are not attached (samples without labels) ⁇ w b
- b 1,..., M ⁇ .
- the unlabeled sample belongs to the sample set B.
- the selection unit 101 of the learning device 100 selects a sample to be labeled from the unlabeled samples stored in the storage unit 107.
- the selection unit 101 calculates g k (w b ; ⁇ ) using Equation (6) for each of the unlabeled samples belonging to the sample set B. That is, the selection unit 101 calculates g k (w b ; ⁇ ) for each of the unlabeled samples w b , where z in Equation (6) is w b .
- the selection unit 101 obtains k for which the expression (6) is the largest and the second largest for each unlabeled sample w b .
- the value of k at which Equation (6) is maximum is i (i is a natural number from 1 to K), and the value of k at which Equation (6) is the second largest is j (j is from 1 to K). Natural number).
- the selection unit 101 sets the i-th class ⁇ i as the first class (first class) and the j-th class ⁇ j as the second class (second class).
- the selection unit 101 calculates r ij for each unlabeled sample w b using the following equation (7).
- This formula (7) is obtained by setting x n in the above-described formula (3) to w b .
- r ij calculated by equation (7) is always negative.
- the larger the value of r ij that is, the closer the value of r ij is to 0, the more the difference between the first class and the second class of the unlabeled sample w b used for calculating r ij.
- the selection unit 101 selects the unlabeled sample w b having the largest value of r ij among the unlabeled samples belonging to the sample set B as a sample to be labeled (referred to as a target sample) (Ste S1).
- the selection unit 101 can preferentially select an unlabeled sample that can be easily identified as a class that is not the correct class.
- the acquisition unit 102 acquires information regarding the labeled sample (labeled sample xn ) (step S2). ). Then, the acquisition unit 102 updates the information associated with the labeled sample stored in the storage unit 107. That is, acquisition unit 102, a set of the label chromatic sample x n belongs, to change from sample set B in sample set A, the label assigned to the label chromatic sample x n in association with the label chromatic samples x n The data is stored in the storage unit 107 (step S3).
- the number M of unlabeled samples belonging to the sample set B is decreased by 1, and the number N of labeled samples belonging to the sample set A is increased by one.
- the updating unit 103 uses the sample set A to update the dictionary (classifier parameter ⁇ ). Specifically, the updating unit 103 updates the parameter ⁇ using the steepest descent method represented by the following equation (8) so that the value of the equation (1) becomes small (step S4).
- ⁇ is a real number greater than zero.
- the calculation unit 104 calculates the loss L N of the labeled sample from Expression (1) using the parameter ⁇ updated by the update unit 103 and the labeled sample x n stored in the storage unit 107. (Step S5).
- the comparison unit 105 sets the loss L N calculated by the calculation unit 104 as L new, and compares this L new with L old (step S6). Thereby, the comparison unit 105 can determine when the loss LN starts to decrease in inverse proportion to the number of labeled samples.
- comparison unit 105 stores loss L new . Specifically, the comparison unit 105 substitutes L new for L old (step S7). Then, the process returns to step S1. Thereafter, the selection unit 101 newly selects a sample to be labeled from the unlabeled samples stored in the storage unit 107 (step S1).
- step S6 When L old > L new (YES in step S6), output unit 106 outputs current parameter ⁇ as a dictionary (step S8).
- the learning apparatus 100 ends the learning process.
- calculation unit 104 uses the dictionary updated by the update unit 103 and one or more labeled samples to calculate the ratio to the number of labeled samples as a loss for the entire labeled sample. . This is because the comparison unit 105 uses this loss to determine whether or not to update the dictionary.
- the loss obtained by the equation having the number of labeled samples in the denominator has the property that the loss decreases in inverse proportion to the number of labeled samples when there are no more samples giving a larger loss.
- a general learning apparatus when labeling a correct class by active learning, the stability of the classifier is often evaluated as an end determination. Therefore, a general learning apparatus performs new learning and evaluation, or prepares an evaluation set necessary for that purpose. Therefore, a general learning apparatus often requires a lot of man-hours and processing time for learning.
- the comparison unit 105 determines the end of the dictionary update using the loss calculated by the calculation unit 104. Therefore, the learning apparatus 100 according to the present embodiment does not need to perform new learning or evaluation for determining the end of the task of labeling the correct class in active learning. Moreover, the learning apparatus 100 according to the present embodiment does not need to prepare a separate evaluation set necessary for that purpose. Therefore, the learning apparatus 100 according to the present embodiment can accurately determine the end of the correct class labeling work in active learning.
- the learning device 100 can accurately determine the end of the labeling work of the correct class, and thus can further improve the man-hours and processing time for learning.
- the calculation unit 104 receives the loss L N, 1 or more loss including the loss L N (L N, L N -1, ⁇ , the L N-h (h is smaller than N An average of an arbitrary natural number)) may be calculated, and the average may be L new .
- the comparison unit 105 calculates a loss (L Nh-1 , L Nh-2 ,..., L N that has not been used when calculating the L new among the previously calculated losses. -Hp (p is an arbitrary natural number smaller than Nh)) may be calculated, and the average may be L old . Then, the comparison unit 105 may compare the L old and the L new .
- the learning device 100 according to the present modification can output a dictionary with higher identification accuracy than the dictionary output by the learning device 100 according to the first embodiment.
- the comparison unit 105 of the learning apparatus 100 compares the loss for the N labeled samples with the previously calculated loss for the (N ⁇ 1) labeled samples. .
- the loss compared by the comparison unit 105 is not limited to this.
- another example of the operation of the comparison unit 105 will be described.
- members having the same functions as those included in the drawings described in the first embodiment described above are given the same reference numerals, and descriptions thereof are omitted.
- FIG. 6 is a functional block diagram showing an example of the functional configuration of the learning apparatus 200 according to the present embodiment.
- the learning device 200 includes a selection unit 101, an acquisition unit 102, an update unit 103, a calculation unit 104, a comparison unit (determination unit) 205, and an output unit 106. And a storage unit 107.
- the comparison unit 205 receives the loss LN of the labeled sample from the calculation unit 104. Then, the received loss L N and the loss L N ⁇ 1 ,..., L Nc (c is stored in the comparison unit 205 and / or the storage unit 107 and previously calculated by the calculation unit 104.
- X q and Y q are calculated by the following equation (9) using an arbitrary natural number smaller than N).
- q is a natural number from 1 to c.
- the comparison unit 205 using the calculated X q and the Y q, the correlation coefficient between X q and Y q, is calculated using the following equation (10).
- the correlation coefficient R is more suitable as an index for determining the end of the task of labeling the correct class.
- the comparison unit 205 compares R with a predetermined threshold, and outputs an instruction to output a dictionary to the output unit 106 when R is larger than the predetermined threshold. In addition, when R is equal to or less than a predetermined threshold, the comparison unit 205 outputs an instruction to select a sample to be labeled to the selection unit 101.
- the predetermined threshold value may be set in advance by the user or may be set by learning.
- the predetermined threshold value may be set to a value more suitable for determining when the loss L N starts to decrease in inverse proportion to the number of labeled samples. Thereby, it can be said that the comparison unit 205 uses the correlation coefficient R to determine when the loss LN starts to decrease in inverse proportion to the number of labeled samples.
- FIG. 7 is a flowchart showing an example of the processing flow of the learning apparatus 200 according to the present embodiment.
- the identification function of the class omega k is assumed to be the formula (6) described above. Further, as a precondition for executing the processing described below, it is assumed that the parameter ⁇ to be set is the same as that in the first embodiment. Similarly to the first embodiment, it is assumed that the storage unit 107 stores N labeled samples and M unlabeled samples.
- steps S11 to S15 shown in FIG. 7 are the same as steps S1 to S5 described above, description thereof will be omitted.
- the comparison unit 205 receives the loss L N calculated by the calculation unit 104, and calculates the correlation coefficient R between X q and Y q using Equation (9) and Equation (10). (Step S16). Then, the comparison unit 205 compares the calculated correlation coefficient R with a predetermined threshold (step S17). Thereby, the comparison unit 205 can determine when the loss LN starts to decrease in inverse proportion to the number of labeled samples.
- comparison unit 205 stores loss LN calculated in step S15. Specifically, the comparing unit 205, then in calculating the correlation coefficient R, so that it can be used this loss L N, and stores the loss L N to the comparator 205 and / or in the storage unit 107 (Step S18). Then, the process returns to step S1. Thereafter, the selection unit 101 newly selects a sample to be labeled from the unlabeled samples stored in the storage unit 107 (step S11).
- the output unit 106 outputs the current parameter ⁇ as a dictionary (step S19).
- the learning apparatus 200 ends the learning process.
- the learning device 200 according to the present embodiment determines the time point at which the loss L N starts to decrease in inverse proportion to the number of labeled samples, similarly to the learning device 100 according to the first embodiment. be able to. Therefore, the learning device 200 according to the present embodiment can have the same effects as the learning device 100 according to the first embodiment.
- FIG. 8 is a functional block diagram illustrating an example of a functional configuration of the learning apparatus 300 according to the present embodiment.
- the learning device 300 includes an update unit 303, a calculation unit 304, and a determination unit 305.
- the learning apparatus 300 may be configured to include the storage unit 107 as in the first embodiment.
- the update unit 303 corresponds to the update unit 103 described above.
- the update unit 303 updates the dictionary used for the classifier based on the determination result of the determination unit 305. Specifically, when the determination unit 305 determines to update the dictionary, the update unit 303 uses a labeled sample to which a new labeled sample (labeled sample) is added, and the loss value is determined. The parameter is changed until convergence, and the dictionary used for the classifier is updated to the parameter (dictionary) at the time when the loss value has converged.
- the calculation unit 304 corresponds to the calculation unit 104 described above.
- the calculation unit 304 calculates the loss for the entire labeled sample using the dictionary updated by the updating unit 303 and one or more samples and labeled samples to which labels have been assigned. Note that the calculation unit 304 calculates a ratio to the number of labeled samples as a loss with respect to the entire labeled sample.
- the determination unit 305 corresponds to the comparison unit 105 or the comparison unit 205 described above.
- the determination unit 305 determines whether to update the dictionary using the loss calculated by the calculation unit 304. If it is determined that the dictionary is to be updated, the determination unit 305 transmits the determination result to the update unit 303. After that, each time the dictionary is updated, the determination unit 305 calculates the loss with respect to the entire labeled sample added with the new labeled sample calculated using the updated dictionary and the updated dictionary. Using the loss for the entire labeled sample before adding a new labeled sample, it is determined whether to update the dictionary, and the determination result is transmitted to the updating unit 303.
- the learning device 300 according to the present embodiment can further improve the man-hours and processing time for learning, similar to the learning device 100 according to the first embodiment.
- FIG. 9 is a diagram illustrating an example of a configuration of a system including learning devices (100, 200, 300) according to each embodiment.
- the system includes a discriminator (identifier) 1 and a learning device (100, 200, 300).
- the discriminator 1 performs discrimination using the dictionary output from the output unit 106 of the learning device (100, 200) or the dictionary when the learning device 300 determines not to update.
- the discriminator can perform discrimination using the dictionary obtained from the learning device (100, 200, 300) according to each embodiment.
- the learning device (100, 200, 300) described above may be realized as a dedicated device, but may be realized using a computer (information processing device).
- FIG. 10 is a diagram illustrating a hardware configuration of a computer (information processing apparatus) capable of realizing each embodiment of the present invention.
- the hardware of the information processing apparatus (computer) 10 shown in FIG. 10 includes the following members.
- CPU Central Processing Unit
- I / F Communication interface
- I / F Communication interface
- I / F Communication interface
- I / F Communication interface
- I / F Communication interface
- ROM Read Only Memory
- -RAM Random Access Memory
- the input / output user interface 13 is a man-machine interface such as a keyboard which is an example of an input device and a display as an output device.
- the communication interface 12 is a general communication means for the devices according to the above-described embodiments (FIGS. 1, 6, and 8) to communicate with an external device via the communication network 20.
- the CPU 11 controls the overall operation of the information processing apparatus 10 that implements the learning apparatus (100, 200, 300) according to each embodiment.
- a program (computer program) that can realize the processing described in each of the above-described embodiments is supplied to the information processing apparatus 10 illustrated in FIG. This can be achieved by reading and executing the above.
- the program is, for example, the various processes described in the flowcharts (FIGS. 5 and 7) referred to in the description of the above embodiments, or the block diagrams shown in FIGS. It may be a program capable of realizing each unit (each block) shown in the apparatus.
- the program supplied to the information processing apparatus 10 may be stored in a readable / writable temporary storage memory (15) or a non-volatile storage device (17) such as a hard disk drive. That is, in the storage device 17, the program group 17 ⁇ / b> A is a program that can realize the functions of the respective units shown in the learning devices (100, 200, 300) in the above-described embodiments, for example. Further, the various kinds of stored information 17B are, for example, learning data, parameters, loss, and the like in the above-described embodiments. However, when the program is installed in the information processing apparatus 10, the constituent unit of each program module is not limited to the division of each block shown in the block diagrams (FIG. 1, FIG. 6, and FIG. 8). May be selected as appropriate during mounting.
- a method for supplying a program into the apparatus can employ a general procedure as follows.
- CD Compact Disc
- a method of installing in the apparatus via various computer-readable recording media (19) such as ROM and flash memory A method of downloading from the outside via a communication line (20) such as the Internet.
- each embodiment of the present invention can be considered to be configured by a code (program group 17A) constituting the computer program or a storage medium (19) in which the code is stored. .
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Machine Translation (AREA)
Abstract
Description
本発明の第1の実施の形態について、図面を参照して詳細に説明する。図1は、本実施の形態に係る学習装置100の機能構成の一例を示す機能ブロック図である。なお、図面中の矢印の方向は、一例を示すものであり、ブロック間の信号の向きを限定するものではない。以降に参照する、他のブロック図においても同様に、図面中の矢印の方向は、一例を示すものであり、ブロック間の信号の向きを限定するものではない。図1に示す通り、本実施の形態に係る学習装置100は、選択部101と、取得部102と、更新部103と、算出部104と、比較部(判定部)105と、出力部106と、記憶部107とを備えている。
算出部104は、損失LNの算出に用いるラベル有サンプルの数Nを、記憶部107に格納されたラベル有サンプルの数を数えることにより、取得してもよいし、取得部102から更新部103を介して取得するものであってもよい。
次に、本実施の形態に係る学習装置100の処理の流れについて説明する。図5は、本実施の形態に係る学習装置100の処理の流れの一例を示すフローチャートである。
本実施の形態に係る学習装置100によれば、学習のための工数および処理時間をより改善することができる。
本実施の形態に係る学習装置100の比較部105が比較する損失の変形例について説明する。比較部105は、算出部104から、損失LNを受け取ると、該損失LNを含む1以上の損失(LN、LN-1、・・・、LN-h(hはNより小さい任意の自然数))の平均を算出し、該平均をLnewとしてもよい。また、比較部105は、以前に算出された損失のうち、上記Lnewを算出する際に使用していない損失(LN-h-1、LN-h-2、・・・、LN-h-p(pはN-hより小さい任意の自然数))の平均を算出し、該平均をLoldとしてもよい。そして、比較部105は、上記LoldとLnewとを比較してもよい。
次に、本発明の第2の実施の形態について説明する。前述した第1の実施の形態では、学習装置100の比較部105は、N個のラベル有サンプルに対する損失と、以前に算出された(N-1)個のラベル有サンプルに対する損失とを比較した。しかしながら、比較部105が比較する損失はこれに限定されるものではない。本実施の形態では、比較部105の動作の他の例について説明する。なお、説明の便宜上、前述した第1の実施の形態で説明した図面に含まれる部材と同じ機能を有する部材については、同じ符号を付し、その説明を省略する。
次に、本実施の形態に係る学習装置200の処理の流れについて説明する。図7は、本実施の形態に係る学習装置200の処理の流れの一例を示すフローチャートである。
本発明の第3の実施の形態について、図面を参照して説明する。本実施の形態では、本発明の課題を解決する最小の構成について説明を行う。図8は、本実施の形態に係る学習装置300の機能構成の一例を示す機能ブロック図である。
次に、図9を参照して、上述した各実施の形態に係る学習装置(100、200、300)を含むシステムの構成について説明する。図9は、各実施の形態に係る学習装置(100、200、300)を含むシステムの構成の一例を示す図である。
ここで、上述した各実施の形態に係る学習装置(100、200、300)を実現可能なハードウェアの構成例について説明する。上述した学習装置(100、200、300)は、専用の装置として実現してもよいが、コンピュータ(情報処理装置)を用いて実現してもよい。
・CPU(Central Processing Unit)11、
・通信インタフェース(I/F)12、入出力ユーザインタフェース13、
・ROM(Read Only Memory)14、
・RAM(Random Access Memory)15、
・記憶装置17、及び
・コンピュータ読み取り可能な記憶媒体19のドライブ装置18。
また、これらはバス16を介して接続されている。入出力ユーザインタフェース13は、入力デバイスの一例であるキーボードや、出力デバイスとしてのディスプレイ等のマンマシンインタフェースである。通信インタフェース12は、上述した各実施の形態に係る装置(図1、図6および図8)が、外部装置と、通信ネットワーク20を介して通信するための一般的な通信手段である。係るハードウェア構成において、CPU11は、各実施の形態に係る学習装置(100、200、300)を実現する情報処理装置10について、全体の動作を司る。
・CD(Compact Disc)-ROM、フラッシュメモリ等のコンピュータ読み取り可能な各種の記録媒体(19)を介して当該装置内にインストールする方法、
・インターネット等の通信回線(20)を介して外部よりダウンロードする方法。
そして、このような場合において、本発明の各実施の形態は、係るコンピュータプログラムを構成するコード(プログラム群17A)或いは係るコードが格納された記憶媒体(19)によって構成されると捉えることができる。
101 選択部
102 取得部
103 更新部
104 算出部
105 比較部
106 出力部
107 記憶部
200 学習装置
205 比較部
300 学習装置
303 更新部
304 算出部
305 判定部
1 識別器
Claims (10)
- 識別器に用いる辞書を更新する更新手段と、
前記更新手段によって更新された辞書と、1以上のサンプルであって、ラベルが付与されたラベル付きサンプルとを用いて、前記ラベル付きサンプルの数に対する比を、該ラベル付きサンプル全体に対する損失として算出する算出手段と、
前記損失を用いて、前記辞書の更新を行うか否かの判定を行う判定手段と、を備え、
前記更新手段は、前記判定手段によって前記辞書の更新を行うと判定された場合、新たなラベル付きサンプルを加えた前記ラベル付きサンプルを用いて、前記辞書を更新し、
前記判定手段は、前記更新された辞書を用いて算出された損失と、更新前の前記辞書を用いて前記新たなラベル付きサンプルを加える前の前記ラベル付きサンプル全体に対して算出された損失と、を用いて、前記辞書の更新を行うか否かの判定を行う、ことを特徴とする学習装置。 - 前記判定手段は、前記更新された辞書を用いて算出された損失が、前記ラベル付きサンプルの数に反比例して減少している場合に、前記辞書の更新を行わないと判定する、ことを特徴とする請求項1に記載の学習装置。
- 前記判定手段は、前記更新された辞書を用いて算出された損失が、前記ラベル付きサンプルの数が一つ少ない数のときに前記算出手段が算出した過去の損失より小さい場合に、前記辞書の更新を行わないと判定する、ことを特徴とする請求項2に記載の学習装置。
- 前記判定手段は、前記更新された辞書を用いて算出された損失および所定数の過去の損失の平均が、前記所定数の過去の損失を算出するより更に前に、前記算出手段が算出した所定数の損失の平均より小さい場合に、前記辞書の更新を行わないと判定する、ことを特徴とする請求項2に記載の学習装置。
- 前記判定手段は、前記ラベル付きサンプルの数の、該ラベル付きサンプルの数が所定数少ない第1のサンプル数に対する比と、前記ラベル付きサンプル全体に対する損失に対する、前記ラベル付きサンプルの数が前記第1のサンプル数の際の損失の比との、相関関数を算出し、前記相関関数が所定の閾値より大きい場合、前記辞書の更新を行わないと判定する、ことを特徴とする請求項2に記載の学習装置。
- ラベルが付与されていないサンプルのうち、正解クラスではないクラスに判別されやすいサンプルを、ラベルを付与する対象のサンプルとして選択する選択手段と、
前記選択手段によって選択された前記ラベルを付与する対象のサンプルに対してラベルが付与されると、前記ラベルを付与する対象のサンプルを含む、前記ラベル付きサンプルを取得する取得手段と、を更に備え、
前記更新手段は、前記取得手段が取得した前記ラベル付きサンプルを用いて、前記辞書を更新する、ことを特徴とする請求項1から5の何れか1項に記載の学習装置。 - 前記判定手段が、前記辞書の更新を行わないと判定したとき、該辞書を出力する出力手段を更に備えることを特徴とする、請求項1から6の何れか1項に記載の学習装置。
- 請求項1から7の何れか1項に記載の学習装置において更新を行わないと判定された前記辞書を用いて、データの識別を行うことを特徴とする識別器。
- 識別器に用いる辞書を更新し、
前記更新された辞書と、1以上のサンプルであって、ラベルが付与されたラベル付きサンプルとを用いて、前記ラベル付きサンプルの数に対する比を、該ラベル付きサンプル全体に対する損失として算出し、
前記損失を用いて、前記辞書の更新を行うか否かの判定を行い、
前記辞書の更新を行うと判定された場合に、新たな前記ラベル付きサンプルを加えたラベル付きサンプルを用いて、前記辞書を更新し、
前記更新された辞書と、前記新たなラベル付きサンプルを加えた前記ラベル付きサンプルとを用いて、前記ラベル付きサンプルの数に対する比を、該ラベル付きサンプル全体に対する損失として算出し、
前記更新された辞書を用いて算出された損失と、更新前の前記辞書を用いて前記新たなラベル付きサンプルを加える前の前記ラベル付きサンプル全体に対して算出された損失と、を用いて、前記辞書の更新を行うか否かの判定を行う、ことを特徴とする学習方法。 - 識別器に用いる辞書を更新する処理と、
前記更新された辞書と、1以上のサンプルであって、ラベルが付与されたラベル付きサンプルとを用いて、前記ラベル付きサンプルの数に対する比を、該ラベル付きサンプル全体に対する損失として算出する処理と、
前記損失を用いて、前記辞書の更新を行うか否かの判定を行う処理と、をコンピュータに実行させ、
前記辞書を更新する処理は、前記辞書の更新を行うと判定された場合、新たな前記ラベル付きサンプルを加えた前記ラベル付きサンプルを用いて、前記辞書を更新する処理であり、
前記辞書の更新を行うか否かの判定を行う処理は、前記更新された辞書を用いて算出された損失と、更新前の前記辞書を用いて前記新たなラベル付きサンプルを加える前の前記ラベル付きサンプル全体に対して算出された損失と、を用いて、前記辞書の更新を行うか否かの判定を行う処理であるプログラムを記憶する、コンピュータ読み取り可能な記録媒体。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2016568362A JP6662301B2 (ja) | 2015-01-07 | 2016-01-05 | 学習装置、識別器、学習方法およびプログラム |
US15/536,783 US11631025B2 (en) | 2015-01-07 | 2016-01-05 | Learning apparatus, learning method, and recording medium |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2015-001271 | 2015-01-07 | ||
JP2015001271 | 2015-01-07 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2016111241A1 true WO2016111241A1 (ja) | 2016-07-14 |
Family
ID=56355934
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2016/000015 WO2016111241A1 (ja) | 2015-01-07 | 2016-01-05 | 学習装置、識別器、学習方法および記録媒体 |
Country Status (3)
Country | Link |
---|---|
US (1) | US11631025B2 (ja) |
JP (1) | JP6662301B2 (ja) |
WO (1) | WO2016111241A1 (ja) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2020187408A (ja) * | 2019-05-10 | 2020-11-19 | 株式会社日立システムズ | 学習データ作成支援システムおよび学習データ作成支援方法 |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US12050866B2 (en) * | 2020-12-13 | 2024-07-30 | International Business Machines Corporation | Maintenance of a data glossary |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2008047835A1 (fr) * | 2006-10-19 | 2008-04-24 | Nec Corporation | Système, procédé et programme d'étude active |
WO2014136316A1 (ja) * | 2013-03-04 | 2014-09-12 | 日本電気株式会社 | 情報処理装置、情報処理方法、及びプログラム |
US20140279716A1 (en) * | 2013-03-15 | 2014-09-18 | Gordon Villy Cormack | Systems and methods for classifying electronic information using advanced active learning techniques |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060112028A1 (en) * | 2004-11-24 | 2006-05-25 | Weimin Xiao | Neural Network and Method of Training |
US7925505B2 (en) * | 2007-04-10 | 2011-04-12 | Microsoft Corporation | Adaptation of language models and context free grammar in speech recognition |
US8219511B2 (en) * | 2009-02-24 | 2012-07-10 | Microsoft Corporation | Unbiased active learning |
US8499008B2 (en) * | 2009-07-24 | 2013-07-30 | Yahoo! Inc. | Mixing knowledge sources with auto learning for improved entity extraction |
JP2011203991A (ja) | 2010-03-25 | 2011-10-13 | Sony Corp | 情報処理装置、情報処理方法、およびプログラム |
TWI648609B (zh) * | 2013-06-07 | 2019-01-21 | 美商科學設計股份有限公司 | 程序監控系統及方法 |
-
2016
- 2016-01-05 US US15/536,783 patent/US11631025B2/en active Active
- 2016-01-05 WO PCT/JP2016/000015 patent/WO2016111241A1/ja active Application Filing
- 2016-01-05 JP JP2016568362A patent/JP6662301B2/ja active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2008047835A1 (fr) * | 2006-10-19 | 2008-04-24 | Nec Corporation | Système, procédé et programme d'étude active |
WO2014136316A1 (ja) * | 2013-03-04 | 2014-09-12 | 日本電気株式会社 | 情報処理装置、情報処理方法、及びプログラム |
US20140279716A1 (en) * | 2013-03-15 | 2014-09-18 | Gordon Villy Cormack | Systems and methods for classifying electronic information using advanced active learning techniques |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2020187408A (ja) * | 2019-05-10 | 2020-11-19 | 株式会社日立システムズ | 学習データ作成支援システムおよび学習データ作成支援方法 |
JP7213138B2 (ja) | 2019-05-10 | 2023-01-26 | 株式会社日立システムズ | 学習データ作成支援システムおよび学習データ作成支援方法 |
Also Published As
Publication number | Publication date |
---|---|
JP6662301B2 (ja) | 2020-03-11 |
US20170344908A1 (en) | 2017-11-30 |
JPWO2016111241A1 (ja) | 2017-10-12 |
US11631025B2 (en) | 2023-04-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110689879B (zh) | 端到端语音转写模型的训练方法、系统、装置 | |
WO2019151507A1 (ja) | 学習装置、学習方法及び学習プログラム | |
US10339471B2 (en) | Ensemble based labeling | |
CN111402865A (zh) | 语音识别训练数据的生成方法、语音识别模型的训练方法 | |
CN106663421B (zh) | 声音识别系统以及声音识别方法 | |
WO2016111241A1 (ja) | 学習装置、識別器、学習方法および記録媒体 | |
US11145308B2 (en) | Symbol sequence estimation in speech | |
US10026022B2 (en) | Face recognition in big data ecosystem using multiple recognition models | |
US10108513B2 (en) | Transferring failure samples using conditional models for machine condition monitoring | |
EP0525640A2 (en) | Dynamic programming matching system for speech recognition | |
JP4533160B2 (ja) | 識別的学習方法、装置、プログラム、識別的学習プログラムを記録した記録媒体 | |
JP5980142B2 (ja) | 学習データ選択装置、識別的音声認識精度推定装置、学習データ選択方法、識別的音声認識精度推定方法、プログラム | |
CN111883109B (zh) | 语音信息处理及验证模型训练方法、装置、设备及介质 | |
KR20090065102A (ko) | 어휘 디코딩 방법 및 장치 | |
CN109710927B (zh) | 命名实体的识别方法、装置、可读存储介质及电子设备 | |
CN110334244B (zh) | 一种数据处理的方法、装置及电子设备 | |
EP3940626A1 (en) | Information processing method and information processing system | |
CN113986245A (zh) | 基于halo平台的目标代码生成方法、装置、设备及介质 | |
CN113807221A (zh) | 轨迹质量识别方法、装置、电子设备及可读存储介质 | |
JP6735956B1 (ja) | モデルパラメタ推定装置、状態推定システムおよびモデルパラメタ推定方法 | |
CN114898201B (zh) | 一种目标检测方法、装置、设备及介质 | |
JP7388239B2 (ja) | 照合装置、照合方法、および、照合プログラム | |
Brejová et al. | A better method for length distribution modeling in HMMs and its application to gene finding | |
CN113571085A (zh) | 语音分离方法、系统、装置和存储介质 | |
CN116959437A (zh) | 语音识别方法、装置、设备、存储介质和程序产品 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 16734983 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2016568362 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 15536783 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 16734983 Country of ref document: EP Kind code of ref document: A1 |