WO2005048184A1 - 能動学習方法およびシステム - Google Patents
能動学習方法およびシステム Download PDFInfo
- Publication number
- WO2005048184A1 WO2005048184A1 PCT/JP2004/014917 JP2004014917W WO2005048184A1 WO 2005048184 A1 WO2005048184 A1 WO 2005048184A1 JP 2004014917 W JP2004014917 W JP 2004014917W WO 2005048184 A1 WO2005048184 A1 WO 2005048184A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- learning
- unknown
- label value
- learned
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 68
- 238000005070 sampling Methods 0.000 claims abstract description 63
- 230000010354 integration Effects 0.000 claims description 52
- 230000006870 function Effects 0.000 claims description 7
- 230000007704 transition Effects 0.000 description 10
- 238000010586 diagram Methods 0.000 description 9
- 230000000694 effects Effects 0.000 description 9
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000007796 conventional method Methods 0.000 description 6
- 230000007246 mechanism Effects 0.000 description 5
- 230000005484 gravity Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 241001655798 Taku Species 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Definitions
- the present invention relates to an active learning method and an active learning system.
- an active learning system For example, consider a learning system that performs statistical analysis on collected data and predicts the results of data with unknown label values based on past data trends. An active learning system can be applied to such a learning system. The outline of this type of active learning system will be described below.
- the data is described as follows.
- One data is described by multiple attributes and labels.
- one of the famous evaluation data is "golf". It determines whether or not you have the power to play golf. It is described from four things: weather, temperature, humidity, and wind strength.
- the weather can be “sunny”, “cloudy” or “rainy”
- the wind takes a value of "Yes” or “No”.
- Temperature and humidity are real numbers. For example, one piece of data might say: weather: sunny, temperature: 15 degrees, humidity: 40%, wind: nothing, play: yes.
- the four attributes of weather, temperature, humidity, and wind are called attributes.
- the result of pre- or non-presence is called a label.
- a value that can be taken by a label is a discrete value, it is particularly called a class.
- the label is binary. Of the two values, the label of interest is the positive example, and the other label is the negative example. If the label is multi-valued, one label value of interest is taken as a positive example, and all other label values are taken as negative examples. When the values that can be taken by the label are continuous values, the label value is called a positive example when there is a label value near the target value, and it is called a negative example when the label value is outside the value.
- ROC receiver operating characteristic
- the ROC curve is defined as follows.
- the ROC curve will be a diagonal line connecting the origin and (1, 1).
- the hit ratio is defined as follows.
- Horizontal axis number of data with known label values z (label value unknown + number of known data), vertical axis: number of positive cases in data with known label values z total number of positive cases.
- the hit ratio is a diagonal line connecting the origin and (1, 1).
- the limit is a line connecting the origin and (the number of positive cases Z (label value is unknown + number of known data), 1)
- Entropy is defined as follows. Each P—i indicates the probability of being i!
- Entropy — ( p —l * log (P-1) + p_2 * log (P-1 2) H—— hP_n * log (P_ n))
- a conventional active learning system is disclosed in Japanese Patent Laid-Open Publication No. 11-316754 [2].
- the active learning system disclosed in this publication includes a learning stage in which a lower algorithm performs learning, a boosting stage in which learning accuracy is improved by boosting, and a boosting stage in which a plurality of input candidate points are improved.
- the active learning algorithm used in the conventional method has a problem in that, at the stage of selecting data to be input, many similar data are output as input points. The reason for this is that there is no mechanism that can fully use the learning algorithm of the lower learning algorithm of the conventional active learning algorithm.
- an object of the present invention is to improve the accuracy of the active learning method, control the accuracy with the user's intention, and provide a function of extracting interesting data first. It is to provide an active learning method.
- Another object of the present invention is to improve the accuracy of the active learning method, control the accuracy with the intention of the user, and provide a function of extracting interesting data first. It is to provide an active learning system.
- An object of the present invention is to provide a storage device that stores a set of known data and a set of unknown data using data with a known label value as known data and data with an unknown label value as unknown data.
- An active learning method using a plurality of learning machines, wherein a plurality of learning machines perform learning after independently sampling storage device power with respect to known data, and a plurality of learning machines as learning results. Integrating and outputting machine output results, multiple learning machines extracting storage device power unknown data and making predictions, calculating and outputting data to be learned next based on the prediction results And inputting the label value corresponding to the data to be learned next, and deleting the data with the input label value from the set of unknown data and adding the data to the set of known data.
- it is achieved by an active learning method that performs unequal weighting.
- the weighting in the active learning method of the present invention for example, when the number of data is extremely biased, the specific gravity is increased.
- the data distribution of the selected candidate data is further selected taking into account the spatial data distribution.
- Another object of the present invention is to provide a storage device that stores a set of known data and a set of unknown data using data with a known label value as known data and data with an unknown label value as unknown data.
- a plurality of learning machines for learning data and predicting unknown data, a plurality of sampling devices provided for each learning machine, and a plurality of sampling devices for sampling known data and inputting the data to a corresponding learning machine;
- a first integration means for integrating the results of learning performed based on the data, and a second integration means for calculating and outputting data to be learned next from prediction results performed by each learning machine based on unknown data.
- Integration means result input means for inputting label values corresponding to data to be learned next, and data with label values deleted from the set of unknown data and added to the set of known data (1) sampling weighting means for setting weights at the time of sampling for each sampling device, and (2) setting weights used when integrating learning results by the first integrating means. Prediction weighting means, (3) data weighting means for setting weights used when selecting data to be learned next by the second integration means, and (4) groups for performing grouping on known data and unknown data. This is achieved by an active learning system having at least one of the generating means.
- the specific gravity is increased.
- weighting data when sampling learning data (2) weighting data when selecting an input point from input candidate points, and (3) input data
- At least one of the three types of weighting, that is, weighting data when making predictions, is used.
- the learning result is treated equally when predicting data.
- the accuracy can be improved by changing the weight. Control becomes possible, and learning with arbitrary accuracy becomes possible.
- the data to be learned next tends to be spatially aggregated in a certain area.
- the conventional active learning method is used. Can be corrected, and the accuracy rate can be increased as compared with the conventional one.
- FIG. 1 is a block diagram showing a configuration of an active learning system according to a first embodiment of the present invention.
- FIG. 2 is a flowchart showing a process of an active learning method using the system shown in FIG. 1.
- FIG. 3A is a graph showing a hit ratio comparing the learning accuracy in the active learning method by the system shown in FIG. 1 with the conventional active learning method.
- FIG. 3B is a graph showing an ROC curve comparing the learning accuracy of the active learning method by the system shown in FIG. 1 with the conventional active learning method.
- FIG. 3C is a graph showing the transition of the correct answer rate, comparing the learning accuracy of the active learning method using the system shown in FIG. 1 with the conventional active learning method.
- FIG. 4 is a block diagram showing a configuration of an active learning system according to a second embodiment of the present invention.
- FIG. 5 is a graph of an ROC curve comparing the learning accuracy between the active learning method using the system shown in FIG. 4 and the conventional active learning method.
- FIG. 6 is a block diagram showing a configuration of an active learning system according to a third embodiment of the present invention.
- FIG. 7 is a graph of an ROC curve comparing the learning accuracy between the active learning method using the system shown in FIG. 6 and the conventional active learning method.
- FIG. 8 is a block diagram showing a configuration of an active learning system according to a fourth embodiment of the present invention.
- FIG. 9 is a graph showing the transition of the correct answer rate, comparing the learning accuracy between the active learning method using the system shown in FIG. 8 and the conventional active learning method.
- FIG. 10 is a block diagram showing a configuration of an active learning system according to a fifth embodiment of the present invention.
- FIG. 11 is a graph of an ROC curve comparing the learning accuracy of the active learning method using the system shown in FIG. 10 with the conventional active learning method.
- FIG. 12 is a block diagram showing a configuration of an active learning system according to a sixth embodiment of the present invention.
- FIG. 13A is a graph showing a hit ratio comparing the learning accuracy of the active learning method by the system shown in FIG. 12 with that of the conventional active learning method.
- FIG. 13B is a graph showing an ROC curve comparing the learning accuracy of the active learning method using the system shown in FIG. 12 with the conventional active learning method.
- FIG. 14 is a block diagram showing a configuration of an active learning system according to a seventh embodiment of the present invention.
- FIG. 15A is a graph showing a hit ratio comparing the learning accuracy of the active learning method by the system shown in FIG. 14 with the conventional active learning method.
- FIG. 15B is a graph showing an ROC curve comparing the learning accuracy of the active learning method by the system shown in FIG. 14 with the conventional active learning method.
- FIG. 16 is a block diagram showing a configuration of an active learning system according to an eighth embodiment of the present invention.
- FIG. 17 is a flowchart showing a process of an active learning method using the system shown in FIG.
- FIG. 18 is a graph showing transition of correct answer rate, comparing learning accuracy in the active learning method by the system shown in FIG. 16 and the conventional active learning method.
- FIG. 19 is a block diagram showing a configuration of an active learning system according to a ninth embodiment of the present invention.
- the active learning system of the present invention provides (1) weighting data when learning data is sampled, (2) weighting data when selecting input candidate point force, and ( 3) To achieve the above-described object of the present invention by employing at least one of a total of three types of weighting, that is, weighting data when predicting input data. It is. In these weighting, when the number of data is extremely biased, weighting is performed to increase the specific gravity. Various embodiments of the present invention can be considered depending on at what stage weighting is performed.
- the selected candidate data is further selected while taking into account the spatial data distribution.
- the selected candidate data is further selected while taking into account the spatial data distribution.
- the active learning system stores a storage device 101 for storing data whose label value is already known (that is, known data), and samples known data in the storage device 101.
- Sampling weighting device 102 for generating data for weighting at the time
- prediction weighting device 103 for generating data for weighting at the time of prediction, and weighting at the time of selecting data to be learned next.
- Rule integration device 107 that summarizes the learning results from learning machine 106, output device 111 connected to rule integration device 107, and multiple learning machines 10 6, a data integration device 108 for calculating the next data to be learned, an output device 112 connected to the data integration device 108, and an output connection for the next data to be learned.
- the control device 110 summarizes the result input by the result input device 113 as, for example, a tabular form, deletes the corresponding data in the storage device 109, and instead replaces the data to which the result is added. Is stored in the storage device 101.
- the sampling device 105 and the learning machine 106 are provided in a one-to-one relationship. Further, each learning machine 106 is supplied with data from the corresponding sampling device 105, and is supplied with unknown data from the storage device 109.
- the output device 111 connected to the rule integration device 107 outputs the learned rules, and the output device 112 connected to the data integration device 108 outputs the next data to be learned.
- Sampling weighting device 102 generates and supplies weighting data for weighting data at the time of sampling based on known data stored in storage device 101, to each sampling device 105. .
- the prediction weighting device 103 generates and generates weighting data for performing weighting when the learning result for each learning machine 106 is put together in the rule integration device 107 based on the known data stored in the storage device 101. The weighting data is supplied to the rule integration device 107.
- the data weighting device 104 performs weighting based on the known data stored in the storage device 101 when the data integration device 108 selects and outputs the next data to be learned. Data is generated, and the generated weighting data is supplied to the data integration device 108.
- weighting by each of the sampling weighting device 102, the prediction weighting device 103, and the data weighting device 104 will be described.
- weighting by these weighting devices 102 to 104 various types can be used as long as they are not uniform.
- the weighting by the sampling weighting device 102 includes, for example, (1) setting a weight according to a class or a value in known data, and (2) when a label value takes a discrete value. However, independently, the weight is set so that all the data of a certain class and the data power of the other classes are also randomly sampled. (3) When the label values take continuous values, each sampling device 105 , Data around a particular label value Weighting so that the data strength of all data and other label values is also randomly sampled.
- the result output from each learning machine 106 can be calculated by class (when the label value takes a discrete value) or by a numerical interval (when the label value is a discrete value). There is a method of determining the weight for each continuous value).
- weighting in the data weighting device 104 for example, (1) assigning weights according to the degree of variation in which the frequency power for each class is calculated when the label values take discrete values, (2) each learning machine Weights are assigned according to the variance of the values obtained as a result of 106, and (3) weights are assigned according to the entropy calculated for each class when the label values take discrete values.
- assigning weights according to the degree of variation it is possible to exclude that only the places where the degree of variation is the maximum have the maximum weight.
- assigning weights according to variance or entropy it may be possible to exclude that only those places where the variance or entropy is the largest have the largest weight.
- a weight may be assigned to the result itself obtained by each learning machine 106 separately from these weights.
- step 201 data whose label value is known is stored in the storage device 101, and data whose label value is unknown is stored in the storage device 109.
- a set of known data is stored in the storage device 101, and a set of unknown data is stored in the storage device 109.
- the sampling weighting device 102 generates weights (ie, weighted data) based on the data sent from the storage device 101, and certain weights read such weights.
- Each sampling device 105 samples known data in the storage device 101 while performing weighting according to the weight sent from the sampling weighting device 102, and sends the sampled data to the corresponding learning machine 106.
- each learning machine 106 executes learning based on the received data.
- the data is also sent from the storage device 101 to the prediction weighting device 103.
- the prediction weighting device 103 performs weighting (that is, weighting) based on the data sent from the storage device 101.
- the rule integration device 107 Reads some such weights and sends them to the rule integrator 107.
- the rule integration device 107 puts together the learning results while weighting the learning results from each learning machine 106 based on the weighting data. At this time, for each result output by each learning machine 106, the frequency is calculated for each class (when the label value takes a discrete value) or for each section of the numerical value (when the label value takes a continuous value), and the frequency is calculated as described above.
- the weight is multiplied, and the one with the largest value is output as the expected value.
- the rule integration device 107 sends the result obtained by combining the learning results to the output device 111 as a rule.
- each learning machine 106 makes a prediction on the data whose label value is unknown in the storage device 109, and the result is sent to the data integration device 108.
- the data has also been sent from the storage device 101 to the data weighting device 104, and in step 206, the data weighting device 104 uses the data sent from the storage device 101 to obtain a weight (that is, the weighted data). ) Or read such weights and send them to the data integrator 108.
- the data integration device 108 does not perform weighting on the prediction result from each learning machine 106 based on the weighting data.
- the results are summarized, and the data to be learned next is selected. The following methods can be used to select the data to be learned next.
- the frequency of the resultant power output by each learning machine 106 is calculated for each class, and the degree of variation or entropy is calculated based on the frequency. Is calculated, and data is selected in descending order of weight assigned according to the degree of variation or entropy.
- Weights are assigned according to the degree of variation or entropy and the result, respectively. In this case, the frequency of the result output by each learning machine 106 is calculated for each class, and a numerical value indicating the degree of variation or entropy is calculated based on the frequency.
- the weights are assigned according to the variance and the result, respectively, the weights are selected in combination with the weight assigned to the result and the weight assigned to the result.
- the variance of the result output from each learning machine 106 is calculated, and the weight assigned according to the variance and the weight assigned to the result are combined to select data in order of decreasing weight.
- the data integration device 108 sends the result to the output device 112 as data to be learned next.
- step 207 the result (label value) for the data to be learned next is input manually or by a computer via the result input device 113.
- the input result is sent to the control device 110, and the control device 110 deletes the data to which the result is input from the storage device 109, and stores the data in the storage device 101 instead.
- the sampling weighting device 102, the prediction weighting device 103, and the data weighting device 104 perform weighting with a uniform V ⁇ deviation.
- FIGS. 3A to 3C illustrate the effects of the active learning system according to the first embodiment.
- a broken line 301 indicates a hit ratio when the conventional active learning method is used
- a solid line 302 indicates a hit ratio when the active learning system of the present embodiment is used. According to the present embodiment, it is understood that the data of the target class (value) is found at an earlier stage than the conventional active learning method!
- a broken line 303 shows an ROC curve when the conventional active learning method is used
- a solid line 304 shows an ROC curve when the active learning system of the present embodiment is used.
- a broken line 305 shows a transition of the correct answer rate when the conventional active learning method is used
- a broken line 306 shows a transition of the correct answer rate when the active learning system of the present embodiment is used.
- the active learning system shown in FIG. 4 is similar to the active learning system of the first embodiment, but differs from that of the first embodiment in that a prediction weighting device and a data weighting device are not provided. Is different from Since the prediction weighting device and the data weighting device are not provided, the rule integration device 107 treats all the results output from the learning machine 106 equally and outputs the final rule by means such as a majority decision. Will be done. Specifically, the rule integrator 107 outputs the result of each learning machine 106 for each class when the label value takes a discrete value, or for each interval in the numerical value when the label value takes a continuous value. Then, the frequency is calculated, and the one with the largest value S is output as the expected value.
- the output result is treated equally, and the data that is most unclear is output.
- the output power of each learning machine 106 calculates the frequency for each class, and calculates a numerical value indicating the degree of variation based on the frequency. Then, the data to be learned next is selected from the data judged to be in a certain class and the data indicating that the index indicating the degree of variation is at or near the maximum.
- the variance is calculated from the result output by each learning machine 106, and the data near a certain numerical value and the data having the maximum or near variance, Select the data to be learned.
- Result output from each learning machine 106 The variance is also calculated for the force, and data other than a specific class (or ⁇ data near a certain numerical value) and "minimum or close to minimum" data From what to learn next Select the data.
- FIG. 5 shows the effect of the active learning system according to the second embodiment.
- a broken line 307 is an ROC curve representing the learning accuracy when the conventional active learning method is used
- a solid line 308 is the class (value) of interest by the active learning system of the present embodiment.
- This is an ROC curve that shows the learning accuracy when sampling is performed so that a large number of data are selected. According to the present embodiment, it can be seen that higher accuracy than the conventional active learning method can be obtained.
- the active learning system shown in FIG. 6 is the same as the active learning system of the first embodiment, except that a force sampling weighting device and a data weighting device are not provided. Is different from Since the sampling weighting device and the data weighting device are not provided, each of the sampling devices 105 treats all known data equally and performs random sampling. Further, in the data integration device 108, similarly to the case of the second embodiment, the output results are treated equally, and the data that is most unclear is output.
- FIG. 7 shows the effect of the active learning system of the third embodiment.
- a line 309 indicates an ROC curve representing the learning accuracy of the active learning system.
- the results were treated evenly when integrating the learning results, so the ability to construct an active learning system with only a certain degree of accuracy was an advantage.
- the system can be configured with the accuracy shown in, for example, A, B, C, and D in FIG.
- the active learning system shown in FIG. 8 is similar to the active learning system of the first embodiment, except that a force sampling weighting device and a prediction weighting device are not provided. Is different from Since the sampling weighting device and the prediction weighting device are not provided, each of the sampling devices 105 handles all known data equally, and performs random sampling.
- the rule integration device 107 As in the case of the embodiment, all the results output from the learning machine 106 are treated equally, and the final rule is output by means such as majority decision.
- FIG. 9 shows the effect of the active learning system according to the fourth embodiment.
- the broken line 310 shows the transition of the correct answer rate when the conventional active learning method is used
- the broken line 311 shows the transition of the correct answer rate when the active learning system of the present embodiment is used.
- the weights at the time of sampling are weighted so that the next data to be experimented is scattered as much as possible. By using such weights, it is clear that learning is faster than with conventional active learning methods.
- the active learning system shown in FIG. 10 is the same as the active learning system of the first embodiment, but differs from that of the first embodiment in that a data weighting device is not provided. Since the data weighting device is not provided, the data integration device 108 treats the output results equally as in the case of the second embodiment, and outputs the data that is most confusing. Become.
- FIG. 11 shows the effect of the active learning system of the fifth embodiment.
- the broken line 312 shows the ROC curve when the conventional active learning method is used
- the broken line 313 shows the ROC curve when the active learning system of the present embodiment is used.
- weighting is performed so that a certain class (value) becomes heavier at the time of sampling, and weighting is similarly performed so that the weight of the class becomes heavier when selecting data to be learned next. Is going.
- the accuracy of learning is improved, and by changing the weight of the prediction weighting device, as shown in A, B, C, and D in FIG. Learning can be performed with high accuracy.
- the active learning system shown in FIG. 12 is the same as the active learning system of the first embodiment, but differs from that of the first embodiment in that a prediction weighting device is not provided.
- the lack of a predictive weighting device allows the rule integrator 107 to use the second implementation. As in the case of the state, all the results from the learning machine 106 are treated equally, and the final rule is output by means such as a majority decision.
- FIGS. 13A and 13B show the effects of the active learning system according to the sixth embodiment.
- a dashed line 314 shows a hit ratio when the conventional active learning method is used
- a solid line 315 shows a hit ratio when the active learning system of the present embodiment is used.
- a dashed line 316 indicates an ROC curve when the conventional active learning method is used
- a solid line 317 indicates an ROC curve when the active learning system of the present embodiment is used.
- weighting is performed so that a certain class (value) becomes heavier at the time of sampling, and similarly, when selecting data to be learned next, the weight of the class becomes heavier. Weighting. According to the present embodiment, 90% of the target class (value) can be found earlier than the conventional one, and the learning accuracy is improved.
- the active learning system shown in FIG. 14 is the same as the active learning system of the first embodiment, but differs from that of the first embodiment in that a sampling weighting device is not provided. Since no sampling weighting device is provided, each of the sampling devices 105 handles all known data equally and performs random sampling.
- FIG. 15A and FIG. 15B show the effect of the active learning system of the seventh embodiment.
- a broken line 318 indicates a hit ratio when the conventional active learning method is used
- a solid line 319 indicates a hit ratio when the active learning system of the present embodiment is used.
- a dashed line 320 indicates an ROC curve when the active learning system of the present embodiment is used.
- the weight of data of a certain class is made heavier, both when selecting the data to be learned next and when combining the learning results.
- the data of the class with the heavier weight is output earlier, and learning can be performed with an arbitrary accuracy as shown in FIGS.
- FIG. 16 The active learning system shown in FIG. 16 is the same as the active learning system of the first embodiment, except that a group generation device 115 is added, and a data integration device and an output connected to the data integration device. The difference is that the device is replaced by the data integration and selection device 114.
- the data integration and selection device 114 has the functions of the data integration device 108 and the output device 112 in the system of the first embodiment (see FIG. 1) when selecting the next data to be learned.
- the group selection device 114 is for grouping data with a known label value stored in the storage device 101, data with an unknown label value stored in the storage device 109, or both data. .
- step 211 data whose label value is known is stored in the storage device 101, and data whose label value is unknown is stored in the storage device 109.
- step 212 the group generation device 115 performs grouping on the known data in the storage device 101 and the unknown data in the storage device 109. The grouping result is output from the group generation device 115 as group information.
- the sampling weighting device 102 generates a weight (that is, weighted data) based on the data sent from the storage device 101, and certain weights are read from such weights.
- a weight that is, weighted data
- each sampling device 105 samples known data in the storage device 101 while performing weighting according to the weight sent from the sampling weighting device 102, and sends the sampled data to the corresponding learning machine 106.
- each learning machine 106 executes learning based on the received data.
- Data is also sent from the storage device 101 to the prediction weighting device 103, and in step 215, the prediction weighting device 103 performs weighting (ie, weighting) based on the data sent from the storage device 101. Data) and read such weights And sends them to the rule synthesizer 107.
- the rule integration device 107 puts together the learning results while weighting the learning results from each learning machine 106 based on the weighting data.
- the rule integration device 107 sends the result obtained by putting together the learning results to the output device 111 as a rule.
- each learning machine 106 makes a prediction on the data whose label value stored in the storage device 109 is unknown, and the result is sent to the data integration and selection device 114.
- the data is also sent from the storage device 101 to the data weighting device 104, and in step 217, the data weighting device 104 uses the weight (ie, weighting) based on the data sent from the storage device 101. Data) or read such weights and send them to the data integration and selection unit 114. Based on the weighting data and the group information from the group generation unit 115, the data integration and selection unit 114 summarizes these results while weighting the prediction results from each learning machine 106, and selects the data to be learned next. I do. At this time, the data integration and selection device 114 makes the next data to be learned according to the grouping performed by the group generation device 814 so that the mutual data is scattered as much as possible.
- the data integration and selection device 114 makes the next data to be learned according to the grouping performed by the group generation device 814 so that the mutual data is scattered as much as possible.
- step 218 the result (label value) for the data to be learned next is input manually or by a computer via the result input device 113.
- the input result is sent to the control device 110, and the control device 110 deletes the data to which the result is input from the storage device 109, and stores the data in the storage device 101 instead. Thereafter, as in the case of the first embodiment, the above-described processing is repeated, and active learning proceeds.
- FIG. 18 illustrates the effect of the active learning system according to the eighth embodiment.
- a broken line 321 shows a transition of the correct answer rate when the conventional active learning method is used
- a broken line 322 shows a change of the correct answer rate when the active learning system of the first embodiment is used.
- a line 323 shows the transition of the correct answer rate when the active learning system of the present embodiment that selects the data to be learned next based on the group information created by the group generation device 115 is used. You. When selecting the next data to be learned based on the information of the group generated by the group generation device, the data of each other should be as small as possible. By selecting data so that they belong to different groups, the accuracy rate can be increased quickly and in stages.
- the present embodiment may be configured so that some or all of the sampling weighting device 102, the prediction weighting device 103, and the data weighting device 104 are not provided.
- the active learning system shown in FIG. 19 is the same as the active learning system according to the eighth embodiment, except that a data selection device 116 is newly provided, and the first device is used instead of the data integration and selection device.
- a data integration device 108 and an output device 112 similar to those of the embodiment are provided.
- the data selection device 118 selects unknown data to be predicted by each learning machine 106 from the storage device 109 in accordance with the group information from the group generation device 115, and selects the selected unknown data from each learning machine 106. It is sent to
- the group generated by the group generation device 115 is sent to the data selection device 116.
- the unknown data is sent from the storage device 109 to the data selection device 116.
- the data selection device 116 selects the unknown data so as to be scattered in different groups as much as possible, and the selected data is sent to the learning machine 106 for prediction.
- the data integration device 108 applies the weight determined by the data weighting device 904 to select the next data to be learned.
- This active learning system has the same effects as the active learning system of the eighth embodiment.
- the present embodiment may be configured so that some or all of the sampling weighting device 102, the prediction weighting device 103, and the data weighting device 104 are not provided.
- the active learning system described above can also be realized by causing a computer such as a personal computer or a workstation to read a computer program for realizing the active learning system and executing the program.
- the program for active learning is a recording medium such as a magnetic tape or CD-ROM.
- Such computers generally include a CPU, a hard disk drive for storing programs and data, a main memory, input devices such as a keyboard and mouse, a display device such as a CRT or liquid crystal display, and a magnetic tape. It is composed of a reading device that reads recording media such as CDs and CD-ROMs, and a communication interface that serves as an interface with a network.
- the hard disk drive, main memory, input device, display device, reading device, and communication interface are all connected to the CPU.
- a recording medium storing a program for executing active learning is attached to a reading device, a recording medium power program is read out and stored in a hard disk device, or such a program is also downloaded from a network. Then, the program is stored in the hard disk device, and then the CPU executes the program stored in the hard disk device, whereby the above-described active learning is executed.
- the scope of the present invention also includes the above-described programs, recording media storing such programs, and program products including such programs.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Description
Claims
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
GB0611998A GB2423395A (en) | 2003-11-17 | 2004-10-08 | Active learning method and system |
US10/579,336 US7483864B2 (en) | 2003-11-17 | 2004-10-08 | Active learning method and system |
JP2005515402A JPWO2005048184A1 (ja) | 2003-11-17 | 2004-10-08 | 能動学習方法およびシステム |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2003-386702 | 2003-11-17 | ||
JP2003386702 | 2003-11-17 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2005048184A1 true WO2005048184A1 (ja) | 2005-05-26 |
Family
ID=34587401
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2004/014917 WO2005048184A1 (ja) | 2003-11-17 | 2004-10-08 | 能動学習方法およびシステム |
Country Status (4)
Country | Link |
---|---|
US (1) | US7483864B2 (ja) |
JP (2) | JPWO2005048184A1 (ja) |
GB (1) | GB2423395A (ja) |
WO (1) | WO2005048184A1 (ja) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2008117343A (ja) * | 2006-11-08 | 2008-05-22 | Nec Corp | 学習処理装置 |
WO2008072459A1 (ja) * | 2006-12-11 | 2008-06-19 | Nec Corporation | 能動学習システム、能動学習方法、及び能動学習用プログラム |
JP2010231768A (ja) * | 2009-03-27 | 2010-10-14 | Mitsubishi Electric Research Laboratories Inc | マルチクラス分類器をトレーニングする方法 |
JP5287251B2 (ja) * | 2006-11-24 | 2013-09-11 | 日本電気株式会社 | 分子間相互作用予測装置の性能評価システム、方法、及びプログラム |
JP2021012494A (ja) * | 2019-07-05 | 2021-02-04 | 公立大学法人会津大学 | ラベリングプログラム、ラベリング装置及びラベリング方法 |
Families Citing this family (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4645288B2 (ja) * | 2005-04-28 | 2011-03-09 | 日本電気株式会社 | 能動学習方法および能動学習システム |
WO2008047835A1 (fr) * | 2006-10-19 | 2008-04-24 | Nec Corporation | Système, procédé et programme d'étude active |
US20080275775A1 (en) * | 2007-05-04 | 2008-11-06 | Yahoo! Inc. | System and method for using sampling for scheduling advertisements in an online auction |
US8086549B2 (en) * | 2007-11-09 | 2011-12-27 | Microsoft Corporation | Multi-label active learning |
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 |
US8775341B1 (en) | 2010-10-26 | 2014-07-08 | Michael Lamport Commons | Intelligent control with hierarchical stacked neural networks |
US9015093B1 (en) | 2010-10-26 | 2015-04-21 | Michael Lamport Commons | Intelligent control with hierarchical stacked neural networks |
US9269053B2 (en) * | 2011-04-28 | 2016-02-23 | Kroll Ontrack, Inc. | Electronic review of documents |
JP6362893B2 (ja) * | 2014-03-20 | 2018-07-25 | 株式会社東芝 | モデル更新装置及びモデル更新方法 |
CN105468887B (zh) * | 2014-09-10 | 2019-03-15 | 华邦电子股份有限公司 | 数据分析系统以及方法 |
US10387430B2 (en) | 2015-02-26 | 2019-08-20 | International Business Machines Corporation | Geometry-directed active question selection for question answering systems |
US10630560B2 (en) * | 2015-04-02 | 2020-04-21 | Behavox Ltd. | Method and user interfaces for monitoring, interpreting and visualizing communications between users |
US11915113B2 (en) * | 2019-03-07 | 2024-02-27 | Verint Americas Inc. | Distributed system for scalable active learning |
GB202001468D0 (en) * | 2020-02-04 | 2020-03-18 | Tom Tom Navigation B V | Navigation system |
US11501165B2 (en) | 2020-03-04 | 2022-11-15 | International Business Machines Corporation | Contrastive neural network training in an active learning environment |
JP7280921B2 (ja) * | 2021-06-15 | 2023-05-24 | 株式会社日立製作所 | 計算機システム、推論方法、及びプログラム |
Family Cites Families (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6158432A (en) * | 1995-12-08 | 2000-12-12 | Cardiopulmonary Corporation | Ventilator control system and method |
US6463930B2 (en) * | 1995-12-08 | 2002-10-15 | James W. Biondi | System for automatically weaning a patient from a ventilator, and method thereof |
JPH1116754A (ja) | 1997-06-20 | 1999-01-22 | Tokin Corp | 積層コイル素子 |
WO2002095534A2 (en) * | 2001-05-18 | 2002-11-28 | Biowulf Technologies, Llc | Methods for feature selection in a learning machine |
WO2000036426A2 (de) * | 1998-12-16 | 2000-06-22 | Siemens Aktiengesellschaft | Verfahren und anordnung zur vorhersage von messdaten anhand vorgegebener messdaten |
US6895081B1 (en) * | 1999-04-20 | 2005-05-17 | Teradyne, Inc. | Predicting performance of telephone lines for data services |
US6487276B1 (en) * | 1999-09-30 | 2002-11-26 | Teradyne, Inc. | Detecting faults in subscriber telephone lines |
JP2001229026A (ja) | 1999-12-09 | 2001-08-24 | Nec Corp | 知識発見方式 |
JP3855582B2 (ja) | 2000-03-06 | 2006-12-13 | Kddi株式会社 | 出力状態判定機能を有する並列ニューラルネットワーク処理システム |
DE10134229A1 (de) * | 2000-08-17 | 2002-02-28 | Ibm | Verfahren und System zum Ermitteln von Abweichungen in Datentabellen |
MY137183A (en) * | 2001-03-16 | 2009-01-30 | Dow Global Technologies Inc | Method of making interpolymers and products made therefrom |
US7152029B2 (en) * | 2001-07-18 | 2006-12-19 | At&T Corp. | Spoken language understanding that incorporates prior knowledge into boosting |
US7143046B2 (en) * | 2001-12-28 | 2006-11-28 | Lucent Technologies Inc. | System and method for compressing a data table using models |
US7188117B2 (en) * | 2002-05-17 | 2007-03-06 | Xerox Corporation | Systems and methods for authoritativeness grading, estimation and sorting of documents in large heterogeneous document collections |
US7167871B2 (en) * | 2002-05-17 | 2007-01-23 | Xerox Corporation | Systems and methods for authoritativeness grading, estimation and sorting of documents in large heterogeneous document collections |
US7328146B1 (en) * | 2002-05-31 | 2008-02-05 | At&T Corp. | Spoken language understanding that incorporates prior knowledge into boosting |
US7136518B2 (en) * | 2003-04-18 | 2006-11-14 | Medispectra, Inc. | Methods and apparatus for displaying diagnostic data |
US7282723B2 (en) * | 2002-07-09 | 2007-10-16 | Medispectra, Inc. | Methods and apparatus for processing spectral data for use in tissue characterization |
US7309867B2 (en) * | 2003-04-18 | 2007-12-18 | Medispectra, Inc. | Methods and apparatus for characterization of tissue samples |
WO2004071455A2 (en) * | 2003-02-13 | 2004-08-26 | Sun Pharmaceuticals Corporation | In vitro prediction of sunscreen pfa values |
-
2004
- 2004-10-08 GB GB0611998A patent/GB2423395A/en not_active Withdrawn
- 2004-10-08 WO PCT/JP2004/014917 patent/WO2005048184A1/ja active Application Filing
- 2004-10-08 JP JP2005515402A patent/JPWO2005048184A1/ja active Pending
- 2004-10-08 US US10/579,336 patent/US7483864B2/en active Active
-
2008
- 2008-12-26 JP JP2008333791A patent/JP2009104632A/ja active Pending
Non-Patent Citations (1)
Title |
---|
ABE.UMAMIZUKA ET AL.: "Nodo gakushu to hakken kagaku", KYORITSU SHUPPAN CO., LTD., 5 May 2000 (2000-05-05), MORISHITA.MIYANO, pages 64 - 72, XP002987087 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2008117343A (ja) * | 2006-11-08 | 2008-05-22 | Nec Corp | 学習処理装置 |
JP5287251B2 (ja) * | 2006-11-24 | 2013-09-11 | 日本電気株式会社 | 分子間相互作用予測装置の性能評価システム、方法、及びプログラム |
WO2008072459A1 (ja) * | 2006-12-11 | 2008-06-19 | Nec Corporation | 能動学習システム、能動学習方法、及び能動学習用プログラム |
JP5187635B2 (ja) * | 2006-12-11 | 2013-04-24 | 日本電気株式会社 | 能動学習システム、能動学習方法、及び能動学習用プログラム |
EP2096585A4 (en) * | 2006-12-11 | 2017-11-15 | NEC Corporation | Active studying system, active studying method and active studying program |
JP2010231768A (ja) * | 2009-03-27 | 2010-10-14 | Mitsubishi Electric Research Laboratories Inc | マルチクラス分類器をトレーニングする方法 |
JP2021012494A (ja) * | 2019-07-05 | 2021-02-04 | 公立大学法人会津大学 | ラベリングプログラム、ラベリング装置及びラベリング方法 |
JP7333496B2 (ja) | 2019-07-05 | 2023-08-25 | 公立大学法人会津大学 | ラベリングプログラム、ラベリング装置及びラベリング方法 |
Also Published As
Publication number | Publication date |
---|---|
GB0611998D0 (en) | 2006-07-26 |
JP2009104632A (ja) | 2009-05-14 |
JPWO2005048184A1 (ja) | 2007-05-31 |
US7483864B2 (en) | 2009-01-27 |
GB2423395A (en) | 2006-08-23 |
US20070094158A1 (en) | 2007-04-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2005048184A1 (ja) | 能動学習方法およびシステム | |
US8825573B2 (en) | Controlling quarantining and biasing in cataclysms for optimization simulations | |
Wang et al. | Differential Evolution Algorithm with Self‐Adaptive Population Resizing Mechanism | |
Kim et al. | Genetic algorithm to improve SVM based network intrusion detection system | |
CN110168578A (zh) | 具有任务特定路径的多任务神经网络 | |
JP5353443B2 (ja) | データ分類器作成装置、データ分類装置、データ分類器作成方法、データ分類方法、データ分類器作成プログラム、データ分類プログラム | |
US11037061B2 (en) | Adaptive cataclysms in genetic algorithms | |
Pappa et al. | Attribute selection with a multi-objective genetic algorithm | |
US11562294B2 (en) | Apparatus and method for analyzing time-series data based on machine learning | |
JP7232122B2 (ja) | 物性予測装置及び物性予測方法 | |
Bäck et al. | Evolutionary algorithms for parameter optimization—thirty years later | |
Mahanta et al. | Evolutionary data driven modeling and multi objective optimization of noisy data set in blast furnace iron making process | |
CN107169029B (zh) | 一种推荐方法及装置 | |
Huang et al. | Harnessing deep learning for population genetic inference | |
CN106919504B (zh) | 一种基于ga算法的测试数据进化生成方法 | |
Zhao et al. | Improved backtracking search algorithm based on population control factor and optimal learning strategy | |
CN112836794B (zh) | 一种图像神经架构的确定方法、装置、设备及存储介质 | |
Loni et al. | Learning activation functions for sparse neural networks | |
Chen et al. | On balancing neighborhood and global replacement strategies in MOEA/D | |
WO2022162839A1 (ja) | 学習装置、学習方法、及び、記録媒体 | |
Sagawa et al. | Learning variable importance to guide recombination | |
Chen et al. | A hybrid replacement strategy for MOEA/D | |
Podgorelec et al. | A multi-population genetic algorithm for inducing balanced decision trees on telecommunications churn data | |
Chen et al. | Composing Synergistic Macro Actions for Reinforcement Learning Agents | |
US20230281496A1 (en) | Equilibrium solution searching method and information processing apparatus |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BW BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE EG ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NA NI NO NZ OM PG PH PL PT RO RU SC SD SE SG SK SL SY TJ TM TN TR TT TZ UA UG US UZ VC VN YU ZA ZM ZW |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): BW GH GM KE LS MW MZ NA SD SL SZ TZ UG ZM ZW AM AZ BY KG KZ MD RU TJ TM AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LU MC NL PL PT RO SE SI SK TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
WWE | Wipo information: entry into national phase |
Ref document number: 2005515402 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2007094158 Country of ref document: US Ref document number: 10579336 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWW | Wipo information: withdrawn in national office |
Country of ref document: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 0611998.6 Country of ref document: GB Ref document number: 0611998 Country of ref document: GB |
|
122 | Ep: pct application non-entry in european phase | ||
WWP | Wipo information: published in national office |
Ref document number: 10579336 Country of ref document: US |