Meta-learning device and program

Abstract

Provided is a meta-learning device that estimates the presence or absence and direction of a causal relationship between random variables, the meta-learning device comprising: a first learning unit that, in a training phase, meta-learns task common parameters of a machine learning model on the basis of training data including at least one of observations or experiments, thereby obtaining learned task common parameters; and a second learning unit that, in a test phase, fixes the values of the learned task common parameters and learns only the task common parameters of the machine learning model, thereby outputting a posterior probability distribution of an adjacency matrix of a causal graph and a posterior probability distribution of an intervention occurrence variable. The second learning unit stochastically samples the intervention occurrence variable, which is a random variable in which an intervention target has occurred, and uses a score function, which makes it possible to evaluate how well the sampled intervention occurrence variable fits data, to update a parameter of the posterior probability distribution in the machine learning model so as to increase the score function.

Description

Meta-learning device and program

[0001] The present disclosure relates to the estimation of the presence and direction of causal relationships between random variables.

[0002] As an expression of the presence and direction of causal relationships (cause and effect) between random variables, there are causal graphs as shown in FIGS. 6(a) and 6(b). The causal graph has each random variable X1, X2, X3,... as vertices, and the direction of the causal relationship between pairs of each random variable is described by directed edges d1, d2, etc. For example, it can describe the gene regulatory network inside a cell.

[0003] The estimation of a causal graph representing the cause-and-effect relationship between random variables has very important significance in various fields of science such as natural science. The data obtained in such fields includes, as shown in FIG. 6, experimental data obtained by an intervention target r and observational data obtained without any intervention with respect to the "mechanism" that determines the values of the random variables. Intervention refers to, for example, in the industrial field, when considering the measured values of each device as random variables, a situation where an abnormality has occurred in the set value of the device (such as a PC) that affects the measured value, and in the life science field, when considering the functional expression level of each gene as a random variable, a situation where a gene mutation that affects the functional expression level has occurred. It is known that a causal graph can be estimated with high accuracy by detecting the difference between data due to the difference in the presence or absence of intervention from at least one of the observational data and experimental data (hereinafter referred to as "observational / experimental data") obtained under all intervention conditions (including the case where there is no intervention) (Non-Patent Document 1).

[0004] In addition, there are some techniques for estimating a causal graph while utilizing the information of a plurality of observational / experimental data sets. However, most of them assume a situation where the intervention experiment is accurate or a situation where the amount of data is sufficient, and it is difficult to apply them to actual observational / experimental data.

[0005] Furthermore, as a technique that takes into account situations where intervention experiments are inaccurate, for example, the technique described in Non-Patent Document 2 proposes probabilistically estimating both the intervention occurrence variable (the random variable in which the intervention target r occurred) and the causal graph structure in each data set, based on Bayesian inference. Probabilistic estimation makes it possible to visualize the uncertainty of the estimation. However, this technique assumes that each observation / experiment dataset contains a sufficient amount of data, and its main focus is on estimating the intervention occurrence variable and causal graph, such as the random variable X1 shown in Figure 6, in all datasets. Therefore, it cannot be used for the purpose of estimating causal graphs in specific, small datasets, which is often required in actual applications.

[0006] To address these challenges, for example in the field of image classification, there is a technique described in Non-Patent Document 3 that introduces a meta-learning approach. This involves learning plausible model parameters as prior knowledge from each training dataset consisting of images of classification classes such as dogs, cats, and mice, and then performing additional training on a test dataset consisting of only a small amount of data, including images of classification classes such as foxes that are not present in the training data.

[0007] Joris M. Mooij, Sara Magliacane, Tom Claassen. "Joint Causal Inference from Multiple Contexts", JMLR, 21(99), p.1 - 108, 2020.Alexander H¨agele, Jonas Rothfuss, Lars Lorch, Vignesh Ram Somnath, Bernhard Sch¨olkopf, Andreas Krause. "BaCaDI: Bayesian Causal Discovery with Unknown Interventions". AISTATS, p.1411-1436,2023.Hang Wu, Wenqi Shi and May D. Wang. "Developing a novel causal inference algorithm for personalized biomedical causal graph learning using meta machine learning" BMC Medical Informatics and Decision Making, 24(137), 2024.

[0008] However, the technique of Non-Patent Document 3 outputs only one estimation result of a plausible causal graph structure. In FIG. 6, for the convenience of the paper, only two causal graph candidates are shown, but when estimating a causal graph, it is necessary to consider a huge number of causal graph candidates. Therefore, even if a single estimation result is output from a small data set, it is difficult to obtain a statistically reliable result.

[0009] The present disclosure has been made in view of the above circumstances, and an object thereof is to propose a meta-learning technique for probabilistically estimating intervention occurrence variables and a causal graph in a specific small data set by performing meta-learning from a plurality of observation / experimental data sets.

[0010] To achieve the above objective, this disclosure provides a meta-learning device for estimating the presence and direction of causal relationships between random variables, comprising: a first learning unit that, in the training phase, obtains learned task common parameters by meta-learning task common parameters of a machine learning model based on training data including at least one of observations and experiments; and a second learning unit that, in the test phase, fixes the values ​​of the learned task common parameters and, based on test data, learns only the task-specific parameters of the machine learning model to output the posterior probability distribution of the adjacency matrix of a causal graph and the posterior probability distribution of the intervention occurrence variable, wherein the second learning unit probabilistically samples the intervention occurrence variable, which is a random variable in which the intervention target occurred, and uses a score function that can evaluate the goodness of fit of the sampled intervention occurrence variable to the data, and updates the parameters of the posterior probability distribution in the machine learning model so that the score function becomes larger.

[0011] As explained above, this disclosure offers the advantage of proposing a meta-learning technique that probabilistically estimates intervention variables and causal graphs in a specific small number of datasets by performing meta-learning from multiple observational and experimental datasets.

[0012] This is a functional configuration diagram of the meta-learning device of the embodiment. (a) is a conceptual diagram showing training data, and (b) is a conceptual diagram showing test data. (a) is a diagram showing an example of posterior probability for each structure of the causal graph, and (b) is a diagram showing an example of posterior probability of intervention occurrence variables in the first candidate. This is an electrical hardware configuration diagram of the meta-learning device of the embodiment. This is a diagram showing the processing method (algorithm) of the embodiment. (a) and (b) are diagrams showing an example of a causal graph.

[0013] Embodiments of the present invention will be described below with reference to the drawings. However, the present invention is not limited to the embodiments shown below, and various modifications are possible without departing from the technical spirit of the invention. Since the drawings are for conceptual explanation of the present invention, dimensions, ratios, or numbers may be exaggerated or simplified as necessary for ease of understanding.

[0014] [Functional Configuration of the Meta-Learning Device] The meta-learning device 20 is composed of one or more computers. When the meta-learning device 20 is composed of multiple computers, it may be referred to as "meta-learning device" or as "learning system".

[0015] Next, the functional configuration of the meta-learning device 20 will be explained using Figure 1. Figure 1 is a functional configuration diagram of the meta-learning device according to this embodiment.

[0016] As shown in Figure 1, the meta-learning device 20 includes an acquisition unit 30, a first learning unit 31, a second learning unit 32, and an output unit 33. Each of these units is a function implemented by instructions from the processor 101 shown in Figure 2, based on a program. The meta-learning device 20 also has a storage unit constructed by a memory 102 or auxiliary storage device 103, which will be described later, containing a machine learning model (neural network) M, a training data management DB (Database) 21, a hyperparameter storage unit 22, and a test data management DB 23.

[0017] <Memory Unit> (Machine Learning Model) The machine learning model M is a neural network that takes predetermined data as input, which is generated based on an unknown causal graph that shows the presence and direction of causal relationships between random variables, and outputs a posterior probability distribution (posterior probability distribution of the adjacency matrix of the causal graph, or the posterior probability distribution of the intervention occurrence variable (the random variable from which the intervention occurred)) for the predetermined data.

[0018] (Training Data Management DB) The training data management DB 21 contains training datasets D1, D2, ..., D, which are generated based on each of the multiple causal graphs, as shown in Figure 2(a). TThis information is stored in memory. A causal graph, for example, in the industrial field, represents the influence relationships between measurements of various pieces of equipment in a factory, and in the life sciences field, it describes the regulatory network between genes within a cell. Multiple causal graphs include data containing random variables that have been intervened (experimental data), such as equipment with abnormal settings or genes that have mutated, and data that do not contain random variables that have been intervened (observational data). It is sufficient that at least one of the experimental or observational data is included. Hereafter, at least one of the observational data or experimental data will be referred to as "observational / experimental data".

[0019] (Hyperparameter storage unit) The hyperparameter storage unit 22 stores the hyperparameters for the first learning unit 31.

[0020] (Test Data Management DB) The test data management DB 23 stores test data generated by a predetermined causal graph, as shown in Figure 2(b).

[0021] <Functional Configuration> (Acquisition Unit) The acquisition unit 30 acquires training datasets, hyperparameters, and test datasets from input from users, etc., or from external devices of the meta-learning device 20. The acquisition unit 30 then stores multiple training datasets in the training data management unit 21. The acquisition unit 30 also stores hyperparameters in the hyperparameter storage unit 22. Furthermore, the acquisition unit 30 stores test datasets in the test data management DB 23.

[0022] (First Learning Unit) In the training phase, the first learning unit 31 estimates the values ​​of task common parameters using the data stored in the training data management DB 21 and the hyperparameter values ​​stored in the hyperparameter storage unit 22.

[0023] More specifically, the first learning unit 31, in the training phase, obtains the learned task common parameters Θ of the neural network (machine learning model) by meta-learning the task common parameters Θ from a data matrix representing the observation / experiment dataset. In this embodiment, the task common parameters Θ are learned in the form of supervised learning for the machine learning model M by predicting the values ​​while masking some of the values ​​in the data matrix. That is, the first learning unit 31 learns a certain variable X from the data matrix. i Obtain the pre-trained task-common parameter Θ of a neural network for predicting one column representing data from the data in other columns.

[0024] Here, the task-common parameter Θ refers to the parameters of the probabilistic model that represents the posterior probability distribution of the adjacency matrix of the causal graph, and the parameters of the fully connected neural network that predicts the values ​​of the data matrix using the adjacency matrix of the causal graph that follows this posterior probability distribution, excluding the coefficient parameters of the final layer.

[0025] The detailed processing of the first learning unit 31 will be explained later.

[0026] (Second Learning Unit) In the test phase, the second learning unit 32 uses the values ​​of the task common parameter Θ learned (estimated) by the first learning unit 31 and the data stored in the test data management DB 23 to estimate and output the posterior probability distributions of the test data (posterior probability distributions of the adjacency matrix of the causal graph and posterior probability distributions of the intervention occurrence variables).

[0027] More specifically, in the test phase, the second learning unit 32 fixes the values ​​of the task-common parameters Θ learned by the first learning unit 31 and learns only the task-specific parameters Φ of the machine learning model M based on the test dataset. At this time, when the second learning unit 32 fits a portion of the task-specific parameters Φ to each data matrix of the test data, it reduces the number of task-specific parameters by using the learned task-common parameters Θ to model a portion of the task-specific parameters Φ, thereby making the task-specific parameters Φ the output of the machine learning model M which has the learned task-common parameters Θ.

[0028] Furthermore, the second learning unit 32 applies the learned task-common parameter Θ to the coefficient parameters of all connected layers in the machine learning model M except for the final layer, and learns only the coefficient parameters of the final layer in the machine learning model M as task-specific parameters Φ.

[0029] Furthermore, the second learning unit 32 probabilistically samples the intervention occurrence variable (the random variable in which the intervention occurred) and uses a score function that also evaluates the predictive performance of the sampled intervention occurrence variable. The model parameters of the posterior probability distribution in the machine learning model M are updated to increase this score function. This makes it possible to computationally efficiently estimate the posterior probability distribution of the adjacency matrix of the causal graph and the posterior probability distribution of the intervention occurrence variable, even when the sample size of the data for which causal relationships are to be estimated is small (see Figure 2(b)).

[0030] In Figure 3, (a) shows examples of posterior probabilities for each causal graph structure, and (b) shows examples of posterior probabilities of intervention variables in the first candidate. As shown in Figure 3(a), the posterior probability for the first candidate causal graph, where there is a directed edge d12 from random variable X1 to random variable X2 and a directed edge d13 from random variable X1 to random variable X3, is 0.084. The posterior probability for the second candidate causal graph, where there is a directed edge d13 from random variable X1 to random variable X3 and a directed edge d23 from random variable X2 to random variable X3, is 0.054. The third candidate and subsequent candidates are omitted.

[0031] Furthermore, as shown in Figure 3(b), in the first candidate causal graph, the posterior probability of the intervention occurrence variable X1, which is the random variable in which the intervention target r occurred, is 0.7. In the first candidate causal graph, the posterior probability of the intervention occurrence variable X2, which is the random variable in which the intervention target r occurred, is 0.1. In the first candidate causal graph, the posterior probability of the intervention occurrence variable X3, which is the random variable in which the intervention target r occurred, is 0.2. Note that examples other than those mentioned above are omitted for the first candidate causal graph. Also, for the second candidate and subsequent candidates, the posterior probabilities of the intervention occurrence variables are output in the same way as for the first candidate.

[0032] The detailed processing of the second learning unit 32 will be explained later.

[0033] (Output Unit) The output unit 33 outputs to the outside the posterior probability distribution of the adjacency matrix of the causal graph and the posterior probability distribution of the intervention occurrence variable, which were output by the second learning unit 32. External output may be to an external device of the meta-learning device 20 (such as the display device 108 described later, a PC (Personal Computer), or a printer).

[0034] [Hardware Configuration] Figure 4 is a hardware configuration diagram of the meta-learning device according to the embodiment.

[0035] As shown in Figure 4, the meta-learning device 20 includes a processor 101, memory 102, auxiliary storage device 103, connection device 104, communication device 105, and drive device 106. The hardware components of the meta-learning device 20 are interconnected via a bus 107.

[0036] The processor 101 acts as a control unit that controls the entire meta-learning device 20 and has various computing devices such as a CPU (Central Processing Unit). The processor 101 reads various programs into the memory 102 and executes them. The processor 101 may also include a GPU (Graphics Processing Unit).

[0037] Memory 102 has main memory devices such as ROM (Read Only Memory) and RAM (Random Access Memory). The processor 101 and memory 102 form a so-called computer, and the computer realizes various functions by having the processor 101 execute various programs read into memory 102.

[0038] The auxiliary storage device 103 stores various programs and various information used when these programs are executed by the processor 101.

[0039] The connection device 104 is a connection device that connects an external device (for example, a display device 108, an operating device 109) to the meta-learning device 20.

[0040] The communication device 105 is a communication device for sending and receiving various types of information with other devices.

[0041] The drive device 106 is a device for setting the recording medium 106m. The recording medium 106m here includes media that record information optically, electrically, or magnetically, such as CD-ROM (Compact Disc Read-Only Memory), flexible disks, magneto-optical disks, etc. The recording medium 106m may also include semiconductor memory that records information electrically, such as ROM (Read Only Memory) or flash memory.

[0042] The various programs to be installed on the auxiliary storage device 103 are installed, for example, when the distributed recording medium 106m is set in the drive device 106 and the various programs recorded on the recording medium 106m are read by the drive device 106. Alternatively, the various programs to be installed on the auxiliary storage device 103 may be installed by downloading them from the network via the communication device 105.

[0043] [Detailed Processing of the Meta-Learning Device] Next, we will explain meta-learning for causal graph estimation from multiple observation / experimental datasets. Here, the meta-learning device 20 processes d random variables X1, ..., X d From the following data matrix D, which consists of n observations relating to these random variables, we obtain the adjacency matrix A ∈ {0, 1} of the causal graph that represents the presence and direction of causal relationships between these random variables. d×d The posterior probability distribution and the random variables X1, ..., X in this data matrix. d A vector I ∈ {0, 1} represents whether or not intervention was performed on [the target]. d We estimate the posterior probability distribution of [the event].

[0044] In this embodiment, consider the case where the sample size n of the data matrix D for which the causal relationship (cause and effect) between random variables is to be estimated is small. To achieve the estimation of the causal relationship under such circumstances, T observation / experimental data sets are considered, where the data matrices D1, …, D have a common causal graph and different intervention-occurring variables. T It is considered to obtain prior knowledge regarding the posterior probability distribution of the adjacency matrix D of the causal graph and the intervention-occurring variable I using these.

[0045] In this embodiment, to solve the above problems, prior knowledge regarding a machine learning model (neural network) is obtained from a plurality of data sets, and a meta-learning approach for effectively performing prediction even when the number of data sets for which prediction is to be achieved is small is introduced. In meta-learning, the index related to each data matrix is called task t, and the model parameters of the machine learning model are considered to be divided into two: task-common parameters Θ and task-specific parameters Φ. In meta-learning, in the training phase, the task-common parameters Θ are learned based on the data matrices D1, …, D T Based on this, and in the test phase, by fixing the value of the task-common parameters Θ to the model parameter values learned in the training phase and only learning the task-specific parameters Φ, application to a small amount of data is made possible.

[0046] As a well-known existing technique in the field of meta-learning, there is a method called model-agnostic meta-learning (MAML) disclosed in Reference 1. (Reference 1) Chelsea Finn, Pieter Abbeel, Sergey Levine. "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks" ICML, p.1126-1135, 2017. This method is a highly versatile method applicable to any machine learning model and is also applicable to the probability distribution model representing the causal graph and intervention-occurring variables according to this embodiment. However, MAML is for task adaptation, that is, the task-specific parameter ψ (t) ∈ Φ is for each data matrix D tWhile stochastic gradient descent is used to achieve a fit, a problem arises when the number of tasks T in the training phase is large, resulting in extremely high computational complexity. In light of this problem, this embodiment formulates a meta-learning model specifically for causal graph estimation that is computationally efficient.

[0047] In this embodiment, similar to the technique disclosed in Reference 2, we consider probabilistically sampling the adjacency matrix of a causal graph to estimate the posterior probability distribution of the adjacency matrix of the causal graph, evaluating the value of a score function that represents the goodness of fit to the data of the sampled adjacency matrix, and updating the model parameters of the posterior probability distribution so that this value increases. (Reference 2) Bertrand Charpentier, Simon Kibler, Stephan Günnemann. "Differentiable DAG Sampling" ICLR, 2022. However, unlike the technique disclosed in Reference 2, we also consider estimating the posterior probability distribution of the intervention occurrence variable, and therefore probabilistically sampling the intervention occurrence variable as well, and modifying the score function so that the goodness of fit to the data of the sampled intervention occurrence variable can also be evaluated.

[0048] The sampling of the adjacency matrix A of the causal graph is performed in the same manner as the technique disclosed in Reference 2. Specifically, using an upper triangular matrix U and an ordered matrix Π as the adjacency matrix A of a directed acyclic graph, the adjacency matrix A is sampled by probabilistically sampling the upper triangular matrix U and the ordered matrix Π using A as shown below.

[0049] Here, the probability distribution of the upper triangular matrix U is a Gumbel softmax distribution with parameter ψ, and the probability distribution of the ordered matrix Π is a Gumbel top-d distribution with parameter ψ. Both are formulated and sampled in the same manner as the technique disclosed in Reference 2.

[0050] Next, the data matrix D of task t. t Intervention-induced variable I (t) The sampling of is parameter ξ tThis is done by formulating it with a Gumbel softmax distribution, and sampling is performed similarly for tasks t = 1, ..., T, with parameter ξ Ι = [ξ1,…,ξ T Use ].

[0051] Furthermore, the first learning unit 31 and the second learning unit 32 perform the following related processing.

[0052] The adjacency matrix A of the sampled causal graph and the intervention occurrence variable I (t) Data matrix D t To evaluate the predictive performance for the random variable X shown below, these are used. (t) Observed value x (t) We formulate a likelihood model for this.

[0053]

[0054] Specifically, the observed vector x (t) The value of the element x at the j ∈ {1,…,d} j (t) This can be represented by the following model:

[0055] Here, A j represents the j-th column vector of the adjacency matrix A. The following symbols represent the Hadamard product.

[0056] ε j represents noise that follows a standard normal distribution, and f j O The j-th random variable X is j The value x when there is no intervention. j This represents a function that predicts f j Ι The j-th random variable X is j The value x when intervention is performed j This represents a function that predicts the result. (Function f) j O And the function f j Ι Each parameter W O , W Ι Model it with a neural network that has the following characteristics. Data Dt To evaluate the score function that represents the goodness of fit for the given condition, we use the log-likelihood function log P(D t | A, I (t) ,W O ,W Ι You just need to evaluate the value of ).

[0057] Task-specific parameters are defined as Φ = {ξ Ι ,W Ι Let} be the common task parameter, and let Θ = {φ,ψ,W O If so, as described above, these parameters can be machine-learned using the meta-learning method MAML disclosed in Reference 1 so that the value of the log-likelihood increases (i.e., the value of the negative log-likelihood, which is its negative value, decreases). However, there are many task-specific parameters Φ, making it difficult to effectively machine-learn these parameters from a small amount of data, and MAML, which uses stochastic gradient descent for machine-learning these parameters, is computationally very complex. Therefore, in this embodiment, the following two formulation improvements are made.

[0058] (1) First, the task-specific parameter ξ t ∈ ξ Ι We consider modeling using task-common parameters by making the output of a machine learning model g, which is a neural network with task-common parameters η. Here, in order to represent the differences between tasks, the input of the machine learning model g is the data matrix D in task t. t As, ξ t = g(D t Let's consider modeling it as η). This will give us task-specific parameters ξ t Machine learning from small datasets becomes unnecessary, and multiple datasets D1, ..., D T If we use machine learning to obtain the task-common parameter η, we can obtain the task-specific parameter ξ t You will be able to obtain the value.

[0059] Next, task-specific parameter W Ι A machine learning model f of a neural network having j ΙLet's consider a computationally efficient formulation of this. When this neural network is formulated as a fully connected neural network, it can be expressed as follows.

[0060] Here, W l (l = 1,…,L) represents the weight parameter matrix of the l-th layer, and ν represents the activation function. In the fully connected neural network described above, only the final layer, i.e., the L-th layer, has the activation function represented by the coefficient vector w. L Expressed as a linear model using , it looks like this:

[0061] In this embodiment, the parameters W of the fully connected neural network model other than the final layer are used. -L = [W1,…,W L-1 ] is used as a common task parameter, and the coefficient parameter w of the final layer is used. L We will treat only as task-specific parameters. This allows us to determine the optimal parameter w from the properties of the linear model. L It becomes possible to analytically calculate the value of .

[0062] Task-common parameters Θ = {φ,ψ,W} in the machine learning of this embodiment O ,η,W -L The procedure for machine learning the} and the task-specific parameter Φ = {wL} is as follows.

[0063] First, we randomly sample tasks t ∈ {1,…,T} and obtain the adjacency matrix A of the causal graph and the intervention occurrence variable I. (t) We randomly sample the data matrix D of task t. t Support set D, consisting of a small number of samples from the sample set. t S ⊂ D t Using only, the task-specific parameter Φ = {w L The optimal value of} is calculated using the following (Equation 1).

[0064] Here, Z, as shown below, is the hidden feature vector z of the (L-1) layer. i This is a matrix consisting of [the specified elements].

[0065] I is the identity matrix, λ Z This is the regularization parameter. This task-specific parameter value

[0066] The matrix obtained by calculating for each j = 1, …, d.

[0067] And the data matrix D other than the support set t Query set D consisting of samples inside t Q Based on this, we minimize the negative log-likelihood using the following equation (Equation 2).

[0068] Here, Ω represents the sampled adjacency matrix A and the sampled intervention occurrence variable I. (t) This represents the regularization function for .

[0069] Figure 5 shows an overview of the learning algorithm (processing method) for task common parameters in the training phase of this embodiment. In the test phase, the learned task common parameters Θ = {φ,ψ,W}, which are the output result of the algorithm shown in Figure 5, are used. O ,η, W -L Using the values ​​of}, the task-specific parameter Φ = {w L By applying machine learning only to}, it becomes possible to estimate the posterior probability distribution of the adjacency matrix of a causal graph and the posterior probability distribution of the intervention-causing variables from a small amount of data.

[0070] [Main Effects of the Embodiment] As described above, in this embodiment, by meta-learning the task common parameter Θ using the algorithm shown in Figure 5 with T observational / experimental data, it is possible to estimate the posterior probability distribution of the adjacency matrix of the causal graph and the posterior probability distribution of the intervention occurrence variable, even when the sample size n of the predetermined data for which causal relationships are to be estimated is small. In other words, by machine learning the probability distribution of the intervention occurrence variable and the model parameters of the probability distribution of the causal graph as prior knowledge from multiple observational / experimental data, it is possible to propose a meta-learning technique that probabilistically estimates the intervention occurrence variable and the causal graph in a specific small dataset.

[0071] [Supplement] The present invention is not limited to the embodiments described above, and may also have configurations or processes (operations) as shown below.

[0072] (1) The meta-learning device 20 can be realized by a computer and a program, but it is also possible to provide the program by recording it on a (non-temporary) recording medium, or by providing the program via a communication network such as the Internet.

[0073] (2) The hardware processor 101 may be single or multiple.

[0074] 20 Meta-learning device (an example of a learning device) 21 Training data management DB (an example of a training data management unit) 22 Hyperparameter storage unit 23 Test data management DB (an example of a test data management unit) 30 Acquisition unit 31 First learning unit 32 Second learning unit 33 Output unit

Claims

1. A meta-learning device for estimating the presence and direction of causal relationships between random variables, comprising: a first learning unit that, in the training phase, obtains learned task common parameters by meta-learning task common parameters of a machine learning model based on training data including at least one of observations and experiments; and a second learning unit that, in the test phase, fixes the values ​​of the learned task common parameters and, based on test data, learns only the task-specific parameters of the machine learning model to output the posterior probability distribution of the adjacency matrix of a causal graph and the posterior probability distribution of the intervention occurrence variable, wherein the second learning unit probabilistically samples the intervention occurrence variable, which is a random variable in which the intervention target occurred, and uses a score function that enables evaluation of the predictive performance for the data of the sampled intervention occurrence variable to update the parameters of the posterior probability distribution in the machine learning model so that the score function becomes larger.

2. The meta-learning apparatus according to claim 1, wherein the second learning unit applies the learned task-common parameters to the coefficient parameters of all connected layers other than the final layer in the machine learning model, and applies the task-specific parameters to be learned by the second learning unit to the coefficient parameters of the final layer in the machine learning model.

3. A program for a computer to implement the meta-learning device described in claim 1 or 2.

Images