Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N20/00—Machine learning

Definitions

• the disclosed technology relates to a machine learning program, a machine learning method, and a machine learning device.
• a learning device has been proposed that inputs training data for learning a classifier and a causal graph that shows the causal relationships between variables included in the training data.
• This learning device uses the input training data and causal graph to learn a classifier by solving a constrained optimization problem in which the average of the causal effects between specified variables is within a specified range and the variance of the causal effects is equal to or less than a specified value.
• an information processing device that generates learning data for making fair judgments on each piece of input data, for example by artificially increasing data with minority attributes.
• This information processing device holds first learning data used to train a machine learning model, acquires information regarding bias in the learning data, and generates second learning data using data included in the learning data based on the information regarding bias. Then, this information processing device trains the machine learning model using the first learning data and the second learning data.
• a system has been proposed that labels unlabeled data according to the amount of label bias.
• the system samples input data according to the amount of selection bias and the discrepancy between feature rarity, and trains a classifier using the sampled and labeled data and additional unlabeled data.
• a method that integrates active learning and semi-supervised learning has been proposed to train machine learning models using fairness-conscious active learning.
• the method selects the most valuable unlabeled data and sends it to an expert system for labeling.
• the method then establishes connections between the unlabeled data and labeled data, improves the model using the unique information of the unlabeled data, and assigns pseudo-labels to those samples.
• the disclosed technology aims to reduce the processing load of fair active learning.
• the disclosed technology calculates the independence between a prediction result when each of a plurality of unlabeled data is input into a machine learning model and a value of a first attribute of each of the plurality of data. Then, the disclosed technology selects first data from the plurality of data based on the independence, obtains a label for the first data, and performs training of the machine learning model based on the first data and the label.
• the disclosed technology has the effect of reducing the processing load of fair active learning.
• FIG. 1 is a diagram for explaining active learning.
• FIG. 1 is a diagram for explaining conventional fair active learning.
• 1 is a diagram for explaining fair active learning in the present embodiment.
• FIG. FIG. 1 is a functional block diagram of a machine learning device according to an embodiment of the present invention.
• FIG. 11 is a diagram for explaining an example of a degree of improvement in accuracy.
• FIG. 1 is a block diagram showing a schematic configuration of a computer that functions as a machine learning device. 1 is a flowchart illustrating an example of a machine learning process. A diagram showing a comparison between the present method and a comparative method in terms of the prediction accuracy, fairness, and execution time of fair active learning of a machine learning model.
• training a machine learning model means learning the relationship between the data labels (outcome variables) and features (explanatory variables) and identifying parameters that approximate that relationship.
• fairness in machine learning means that the prediction results of a machine learning model are free of bias or discrimination based on the innate or acquired characteristics (hereinafter referred to as "protected attributes") of individuals or groups in decision-making.
• protected attributes innate or acquired characteristics
• fairness and prediction accuracy of a machine learning model it is also important to strike a balance between the two.
• labeled data When training a machine learning model using supervised learning, data with labels indicating the correct answer (hereinafter referred to as "labeled data") is required.
• Labeled data is data that is tagged with one or more labels.
• the labels are usually assigned by humans or other information sources such as oracles. Without a sufficient amount of labeled data, the fairness and predictive accuracy of the machine learning model cannot be sufficiently improved. However, labeled data is more expensive to collect than unlabeled data (hereinafter referred to as “unlabeled data”).
• Active learning is an interactive machine learning method that improves a machine learning model by asking questions. Specifically, as shown in Figure 1, an information processing device that performs active learning performs the following processes: (1) data selection, (2) questions, (3) answers, (4) training, and (5) transmission.
• the information processing device calculates a data acquisition function for each data included in the unlabeled data set as a process of "(1) data selection", and preferentially selects data useful for training the machine learning model based on the data acquisition function.
• the data acquisition function is an index that indicates the ambiguity of the prediction by the machine learning model for each data, the representativeness of each data for the unlabeled data set, etc., using information such as the parameters of the current machine learning model.
• the information processing device queries the oracle for the label of the selected data as a process of "(2) question”.
• the information processing device obtains the label, which is the answer from the oracle, as a process of "(3) answer", and assigns the obtained label to the selected data to make it labeled data, which is added to the labeled data set. Furthermore, the information processing device trains the machine learning model using the labeled data set as a process of "(4) training". Furthermore, the information processing device transmits information such as the parameters of the machine learning model after training to the process of (1) data selection as a process of "(5) transmission".
• the above steps (1) to (5) are repeatedly executed, and from among the unlabeled data, labeling is preferentially performed on data that is useful for training a machine learning model, making it possible to collect labeled data effectively.
• (1) data selection involves selecting data that will help improve the predictive accuracy of the machine learning model, without considering fairness. As a result, as more labeled data is added and the training of the machine learning model progresses, fairness may worsen. For example, in a machine learning model for facial expression recognition, it is conceivable that only data of a certain race may be selected from a set of unlabeled data.
• (1) in the data selection process data is selected taking into consideration the trade-off between fairness and prediction accuracy.
• data is selected taking into consideration the trade-off between fairness and prediction accuracy.
• FIG. 2 in conventional fair active learning, a part of the unlabeled data set is set as an unlabeled validation data set, and the rest is set as an unlabeled candidate data set.
• An information processing device that performs conventional fair active learning estimates the degree of unfairness improvement of the machine learning model in the validation data for each candidate data by executing the processes of (A) tentative question, (B) tentative answer, (C) training, (D) evaluation, and (E) selection shown in FIG. 2.
• the information processing device inputs each candidate data to a labeling model that outputs a temporary label for the data as a process of "(A) temporary question”. Furthermore, as a process of "(B) temporary answer”, the information processing device acquires the temporary label output from the labeling model and assigns the acquired temporary label to each candidate data to create a temporary labeled candidate data set. Furthermore, as a process of "(C) training”, the information processing device trains a machine learning model using the temporary labeled candidate data set. Furthermore, as a process of "(D) evaluation”, the information processing device uses an unlabeled validation data set to evaluate the fairness of each temporary labeled candidate data, taking into account the difference in unfairness of the machine learning model before and after training.
• the information processing device evaluates the prediction accuracy of the machine learning model using the unlabeled validation data set. Furthermore, as a process of "(E) selection", the information processing device selects the candidate data with the best value based on an index that takes into account the trade-off between fairness and prediction accuracy.
• the fairness of each candidate data is evaluated based on the independence between the model's prediction results for unlabeled data and the values of the protection attributes, without the need to train a machine learning model.
• one method that comes easily to mind for solving the above problem is to estimate the degree of unfairness of each candidate data based on the prediction results of the machine learning model for each piece of candidate data, and select data with a low degree of unfairness from the unlabeled candidate data set.
• the data selected does not take into account the impact on the validation data, so outliers or similar data are likely to be selected, resulting in low representativeness of the data. Therefore, in this embodiment, the prediction results of the validation data when candidate data are given are used as the prediction results of the model for unlabeled data.
• the machine learning device will be described below.
• a labeled data set 20 and an unlabeled data set 22 are input to the machine learning device 10.
• the amount of labeled data included in the labeled data set 20 is much smaller than the amount of unlabeled data included in the unlabeled data set 22.
• the machine learning device 10 selects data from the unlabeled data set 22 with consideration given to fairness, and performs training of the machine learning model 24. In other words, the machine learning device 10 performs fair active learning.
• the machine learning device 10 functionally includes a control unit 11, as shown in FIG. 4.
• the control unit 11 further includes a training unit 12, a calculation unit 14, a selection unit 16, and an acquisition unit 18.
• a machine learning model 24 is also stored in a specified storage area of the machine learning device 10.
• the training unit 12 uses a plurality of labeled data included in the labeled data set 20 as training data to train the machine learning model 24. As described below, in this embodiment, new labeled data is added to the labeled data set 20 by the acquisition unit 18. When new labeled data is added to the labeled data set 20, the training unit 12 uses the original labeled data and the added labeled data to train the machine learning model 24.
• the calculation unit 14 calculates the independence between the prediction result (hereinafter referred to as the "predicted label") when each of the multiple unlabeled data included in the unlabeled data set 22 is input to the machine learning model 24 and the value of the protection attribute of each of the multiple data.
• the protection attribute is an example of a "first attribute" of the disclosed technology.
• the calculation unit 14 calculates the mutual information between the predicted label and the value of the protection attribute as the independence.
• the mutual information is an index that quantitatively represents whether or not two variables are dependent, and if the two variables are completely independent of each other, the mutual information is 0. In other words, if the mutual information between the predicted label and the value of the protection attribute is 0, it can be said that the machine learning model 24 is completely fair. Therefore, it can be said that the unlabeled data with the smallest mutual information is the most fair data.
• the calculation unit 14 calculates the degree of unfairness improvement of the machine learning model 24 based on the mutual information between the predicted label and the value of the protected attribute for the validation data, conditioned on each candidate data. Specifically, the calculation unit 14 sets a part of the multiple unlabeled data included in the unlabeled data set 22 as validation data, and sets the data other than the validation data as candidate data. The calculation unit 14 calculates the independence between the predicted label and the value of the protected attribute for each validation data, which is conditioned on the independence between the predicted label and the value of the protected attribute for each candidate data.
• the calculation unit 14 calculates a difference in mutual information I between the predicted label Yv of the verification data v and the value Sv of the protection attribute in the verification data v before and after the candidate data u is provided as the unfairness improvement degree F u of the candidate data u using the following formula (1).
• Y u is the predicted label of candidate data u
• S u is the value of the protection attribute in candidate data u.
• the first term in ⁇ on the right side of formula (1) is the mutual information of the verification data v before the candidate data u is given
• the second term is the mutual information of the verification data v after the candidate data u is given.
• the larger the value of the unfairness improvement degree F u of candidate data u the higher the degree of improvement in unfairness.
• the calculation unit 14 calculates the second term in the ⁇ on the right-hand side of equation (1) as follows. First, the calculation unit 14 converts the second term into equation (2) below. In equation (2), H(X) is the entropy of X.
• the calculation unit 14 approximates a probability distribution corresponding to each entropy by Monte Carlo dropout.
• the calculation unit 14 assumes that the parameters of the machine learning model 24 and the distribution of the predicted labels of the machine learning model 24 are conditionally independent, and calculates the probability P( Yi ) of Yi by the following formula (3).
• Yi ⁇ Yv , Sv , Yu , Su ⁇ .
• Equation (3) ⁇ is a parameter of the machine learning model 24, and M is the number of Monte Carlo samplings.
• the calculation unit 14 calculates the mutual information of equation (2) using the probability distribution of equation (3).
• the calculation unit 14 calculates the degree of improvement in the prediction accuracy of the machine learning model 24 by each candidate data based on the uncertainty of the prediction result when each candidate data is input to the machine learning model 24 for each candidate data.
• data located near the decision boundary of the machine learning model 24 can be said to be data that is difficult for the machine learning model 24 to judge, so by using such data as training data, the prediction accuracy of the machine learning model 24 is improved.
• the decision boundary is defined in the feature space. Note that the example of FIG. 5 shows an example of binary classification of a linear model, and each circle represents the feature amount of each data. In this case, data near the decision boundary (for example, data shown by a dotted circle in FIG.
• the calculation unit 14 calculates the degree of accuracy improvement that is higher as the candidate data is closer to the decision boundary of the machine learning model. For example, the calculation unit 14 calculates the entropy indicating the uncertainty of the predicted label Y u of the candidate data u as the accuracy improvement degree A u , as shown in the following formula (4).
• Y is a set of predicted labels of the machine learning model 24.
• the calculation unit 14 calculates an evaluation value Eu for each candidate data u , which is represented by an unfairness improvement degree F u, an accuracy improvement degree A u , and a coefficient ⁇ representing a trade-off between the unfairness improvement degree F u and the accuracy improvement degree A u , for example, as shown in the following equation ( 5) .
• E u ⁇ F u +(1- ⁇ ) ⁇ A u (5)
• ⁇ is a coefficient that ranges from 0 to 1 (for example, 0.6) and specifies to what degree the unfairness improvement degree F u or the accuracy improvement degree A u is to be prioritized.
• the selection unit 16 selects target data to be labeled from the multiple candidate data u based on the evaluation value Eu for each candidate data u calculated by the calculation unit 14.
• the target data is an example of the "first data" of the disclosed technology.
• the selection unit 16 may select candidate data u with the highest evaluation value Eu, may select candidate data u with an evaluation value Eu equal to or greater than a predetermined value, or may select candidate data u with a top predetermined number of evaluation values Eu.
• the acquisition unit 18 queries an oracle, which is a human or other information source, for the label of the target data selected by the selection unit 16, and acquires the label that is the response from the oracle.
• the acquisition unit 18 assigns the acquired label to the target data to make it labeled data, and adds it to the labeled data set 20.
• the training unit 12 also uses the added labeled data to train the machine learning model 24.
• the machine learning device 10 may be realized, for example, by a computer 40 shown in FIG. 6.
• the computer 40 includes a CPU (Central Processing Unit) 41, a GPU (Graphics Processing Unit) 42, a memory 43 as a temporary storage area, and a non-volatile storage device 44.
• the computer 40 also includes an input/output device 45 such as an input device and a display device, and an R/W (Read/Write) device 46 that controls the reading and writing of data from and to a storage medium 49.
• the computer 40 also includes a communication I/F (Interface) 47 that is connected to a network such as the Internet.
• the CPU 41, GPU 42, memory 43, storage device 44, input/output device 45, R/W device 46, and communication I/F 47 are connected to each other via a bus 48.
• the storage device 44 is, for example, a hard disk drive (HDD), a solid state drive (SSD), flash memory, etc.
• the storage device 44 which serves as a storage medium, stores a machine learning program 50 for causing the computer 40 to function as the machine learning device 10.
• the machine learning program 50 has training process control instructions 52, calculation process control instructions 54, selection process control instructions 56, and acquisition process control instructions 58.
• the storage device 44 also has an information storage area 60 in which information constituting the machine learning model 24 is stored.
• the CPU 41 reads the machine learning program 50 from the storage device 44, expands it in the memory 43, and sequentially executes the control instructions of the machine learning program 50.
• the CPU 41 operates as the training unit 12 shown in FIG. 4 by executing the training process control instruction 52.
• the CPU 41 also operates as the calculation unit 14 shown in FIG. 4 by executing the calculation process control instruction 54.
• the CPU 41 also operates as the selection unit 16 shown in FIG. 4 by executing the selection process control instruction 56.
• the CPU 41 also operates as the acquisition unit 18 shown in FIG. 4 by executing the acquisition process control instruction 58.
• the CPU 41 also reads information from the information storage area 60 and expands the machine learning model 24 in the memory 43.
• the computer 40 that has executed the machine learning program 50 functions as the machine learning device 10.
• the CPU 41 that executes the program is hardware. A part of the program may be executed by the GPU 62.
• the functions realized by the machine learning program 50 may be realized, for example, by a semiconductor integrated circuit, more specifically, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), etc.
• ASIC Application Specific Integrated Circuit
• FPGA Field-Programmable Gate Array
• the machine learning device 10 executes the machine learning process shown in FIG. 7.
• the machine learning process is an example of a machine learning method of the disclosed technology.
• step S10 the training unit 12 acquires the labeled data set 20 and uses the labeled data as training data to train the machine learning model 24.
• step S12 the training unit 12 determines whether or not a termination condition for fair active learning is met.
• the termination condition may be, for example, when the number of data newly added to the labeled data set 20 exceeds a predetermined number. If the termination condition is not met, the process proceeds to step S14.
• step S14 the calculation unit 14 sets a part of the multiple unlabeled data included in the unlabeled data set 22 as validation data and the rest as candidate data.
• step S16 the calculation unit 14 calculates the difference in mutual information I between the predicted label Y v of validation data v and the value S v of the protection attribute in validation data v before and after the candidate data u is provided as the unfairness improvement degree F u of each candidate data u, for example, as shown in formula (1).
• step S18 the calculation unit 14 calculates the entropy indicating the uncertainty of the predicted label Y u of each candidate data u as the accuracy improvement degree A u , for example, as shown in formula (4).
• step S20 the calculation unit 14 calculates an evaluation value E u for each candidate data u, which is represented by the unfairness improvement degree F u , the accuracy improvement degree A u , and a coefficient ⁇ indicating the trade-off between the unfairness improvement degree F u and the accuracy improvement degree A u , for example, as shown in formula ( 5 ).
• step S22 the selection unit 16 selects target data to be labeled from the multiple candidate data u based on the evaluation value Eu for each candidate data u.
• step S24 the acquisition unit 18 inquires of the oracle about the label of the target data and acquires the label that is the response from the oracle.
• step S26 the acquisition unit 18 assigns the acquired label to the target data to make it labeled data and adds it to the labeled data set 20, and also deletes the candidate data that became the target data from the unlabeled data set 22, and returns to step S10.
• step S10 the training unit 12 trains the machine learning model 24 using the added labeled data.
• step S12 the termination condition for fair active learning is met.
• step S28 the training unit 12 outputs the machine learning model trained by fair active learning, and the machine learning process ends.
• the machine learning device calculates the independence between the predicted result when each of the multiple unlabeled data is input to the machine learning model and the value of the protection attribute of each of the multiple unlabeled data. Furthermore, the machine learning device selects target data from the multiple unlabeled data based on the calculated independence, queries the oracle to obtain a label for the target data, and trains the machine learning model based on the target data and the obtained label. In other words, the machine learning device according to this embodiment evaluates the fairness of the unlabeled data based on an information theoretical approach, and selects data to be labeled without re-learning the machine learning model. This allows the machine learning device according to this embodiment to reduce the processing load of fair active learning.
• the machine learning device also calculates the degree of unfairness improvement for each candidate data using mutual information as the independence between the prediction result and the value of the protection attribute, and calculates the degree of improvement in accuracy of the machine learning model based on the uncertainty of the candidate data.
• the machine learning device selects target data based on an evaluation value that takes into account the trade-off between the degree of unfairness improvement and the degree of accuracy improvement. This allows the machine learning device according to this embodiment to select as target data data that optimizes the trade-off between the degree of unfairness improvement and the degree of accuracy improvement while reducing the processing load.
• the machine learning device treats part of the unlabeled data as validation data and the rest as candidate data, and calculates the difference in mutual information between the prediction result of the validation data and the value of the protected attribute before and after providing the candidate data as the degree of unfairness improvement. In this way, by calculating the degree of unfairness improvement of the candidate data taking into account the impact on the validation data, it becomes easier to select data with a high level of representativeness from the candidate data.
• FIG. 8 shows a schematic comparison of the prediction accuracy, fairness, and execution time of fair active learning of the machine learning model between the method of this embodiment (hereinafter referred to as "this method") and the comparative method.
• the comparative method here is a method that requires re-learning of the machine learning model when selecting data to be labeled, as in the method described in Non-Patent Document 1.
• the trade-off between prediction accuracy and fairness of this method is equivalent to that of the comparative method.
• the execution time of this method is significantly reduced compared to the comparative method.
• the theoretical calculation cost of fair active learning is "O((T+N v )CN u )" for the comparative method and "O(N u N v C 2 M)" for this method.
• T is the calculation cost of training the machine learning model
• N v is the number of validation data
• N u is the number of candidate data
• C is the number of labels
• M is the number of Monte Carlo samplings.
• mutual information is used as an index showing the independence between the prediction results of a machine learning model for unlabeled data and the values of protected attributes.
• the mutual information can be anything that can mathematically define the independence between the prediction results and the values of protected attributes, and for example, Kullback-Leibler divergence, Jensen-Shannon divergence, covariance, demographic parity difference, disparate impact ratio, etc. may be applied.
• the machine learning program is pre-stored (installed) in the storage device, but this is not limited to this.
• the program related to the disclosed technology may be provided in a form stored in a storage medium such as a CD-ROM, DVD-ROM, or USB memory.
• Machine learning device 11
• Training unit 14
• Calculation unit 16
• Selection unit 18
• Acquisition unit 20
• Labeled data set 22
• Unlabeled data set 24
• Machine learning model 40
• Computer 41
• GPUs 43
• Memory 44
• Input/output device 46
• R/W device 47
• Communication I/F 48
• Bus 49
• Machine learning program 52
• Training process control instructions 54
• Selection process control instructions 56
• Acquisition process control instructions 60
• Information storage area 60

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Software Systems (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• Medical Informatics (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Physics & Mathematics (AREA)
• Computing Systems (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Mathematical Physics (AREA)
• Artificial Intelligence (AREA)
• Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Priority Applications (4)

Applications Claiming Priority (1)

Related Child Applications (1)

Publications (1)

Family

ID=92262161

Family Applications (1)

Country Status (4)

Citations (5)

• 2023
  • 2023-02-09 WO PCT/JP2023/004458 patent/WO2024166331A1/ja not_active Ceased
  • 2023-02-09 JP JP2024576024A patent/JPWO2024166331A1/ja active Pending
  • 2023-02-09 EP EP23921162.6A patent/EP4664361A4/en active Pending
• 2025
  • 2025-06-16 US US19/238,994 patent/US20250307714A1/en active Pending

Patent Citations (5)

Non-Patent Citations (3)

Also Published As

Similar Documents

Legal Events