JP7647934B2 - DETECTION METHOD, DETECTION APPARATUS, AND PROGRAM - Google Patents

Description

The present invention relates to a detection method, a detection device, and a program.

When predicting classes (or classifying classes) using a machine learning model, predictions for data of classes not included in the training data are often significantly inaccurate. For this reason, out-of-distribution data detection techniques (e.g., Non-Patent Document 1), which are techniques for detecting data of classes not included in the training data, are important. Out-of-distribution data detection techniques make it possible to detect data that may be significantly inaccurate and process that data separately. In addition, out-of-distribution data detection techniques can also detect abnormal data and check the quality of collected data.

Liang, Shiyu, Yixuan Li, and R. Srikant. "Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks." International Conference on Learning Representations. 2018.

However, conventional out-of-distribution data detection methods have the problem that they require large amounts of training data.

One embodiment of the present invention has been made in consideration of the above points, and aims to detect out-of-distribution data with high accuracy.

In order to achieve the above-mentioned objective, a detection method according to one embodiment includes a meta-learning procedure in which a first training data set consisting of training data for a plurality of tasks including a target task, which is a desired task, is used to learn a common model, which is a machine learning model common to the plurality of tasks, and a detection procedure in which a computer executes the meta-learning procedure in which a first training data set consisting of training data for the target task is used to learn a common model, which is a machine learning model common to the plurality of tasks, and a detection procedure in which a computer executes the detection procedure in which a label not assigned to the training data included in the second training data set is used as out-of-distribution data, using the trained common model, a second training data set consisting of training data for the target task, and a target task-specific model, which is a machine learning model for the target task.

Out-of-distribution data can be detected with high accuracy.

FIG. 2 is a diagram illustrating an example of a hardware configuration of a detection device according to the present embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of a detection device according to the present embodiment. 13 is a flowchart illustrating an example of a meta-learning process according to the present embodiment. 10 is a flowchart illustrating an example of a detection process according to the present embodiment.

An embodiment of the present invention will be described below. In this embodiment, a detection device 10 will be described that can detect out-of-distribution data in a target task with high accuracy even when there is only a small amount of training data for the target task. Here, training data refers to data represented as a pair of features of an example and a label indicating which class the features belong to.

The detection device 10 of this embodiment has a meta-learning time in which meta-learning is performed using learning data of multiple tasks including the target task, and a detection time in which, when data (feature) of the target task is given, it is determined whether the feature is out-of-distribution data.

During meta-learning, a model for each task (hereafter also called a task-specific model), a model common to all tasks (hereafter also called a common model), and a learning data set for meta-learning are

とが検出装置１０に与えられるものとする。ここで、Ｔはタスク数、Ｎ_ｔはｔ番目のタスクの学習データ数（言い換えれば、ｔ番目のタスクのラベル付き事例数）、ｘ_ｔｎはｔ番目のタスクのｎ番目の事例の特徴量を表す。また、

and are given to the detection device 10. Here, T is the number of tasks, _Nt is the number of learning data for the t-th task (in other words, the number of labeled examples for the t-th task), and _xtn is the feature amount of the n-th example for the t-th task.

はｘ_ｔｎのラベル（クラスラベル）、ｃ_ｔｋはｔ番目のタスクのｋ番目のクラス、Ｋ_ｔはｔ番目のタスクのクラス数を表す。なお、以下では、Ｄ_ｔ＝｛（ｘ_ｔｎ，ｙ_ｔｎ）｜ｎ＝１，・・・，Ｎ_ｔ｝と表記する。また、ｔ番目のタスクを「タスクｔ」、ｎ番目の事例を「事例ｎ」、ｋ番目のクラスを「クラスｋ」等とも表記する。

represents the label (class label) of _xtn , _ctk represents the kth class of the tth task, and _Kt represents the number of classes of the tth task. In the following, _Dt is written as {( _xtn , _ytn )|n=1,..., _Nt }. The tth task is also written as "task t", the nth case as "case n", the kth class as "class k", etc.

At the time of detection, the common model learned during meta-learning, the task-specific model for the target task t', a training data set D _t' = {(x _t'n , y _t'n )|n=1, ..., N _t' } for the target task t', and a feature x to be determined as being out-of-distribution data (or a set of such features to be determined {x _n |n=1, ..., N}, where N is the number of features) are given. Here, it is assumed that the number of training data N _t' included in the training data set D _t' for the target task t' is relatively small (i.e., smaller than the number of training data required by conventional out-of-distribution data detection methods).

Note that each task-specific model (machine learning model) can be any model depending on the task, but in the following we will assume that it is an arbitrary density distribution model as an example.

<Hardware configuration of detection device 10>
An example of a hardware configuration of the detection device 10 according to the present embodiment is shown in Fig. 1. As shown in Fig. 1, the detection device 10 according to the present embodiment includes an input device 101, a display device 102, an external I/F 103, a communication I/F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these pieces of hardware is connected to each other via a bus 109 so as to be able to communicate with each other.

The input device 101 is, for example, a keyboard, a mouse, a touch panel, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the detection device 10 does not have to have at least one of the input device 101 and the display device 102, for example.

The external I/F 103 is an interface with an external device such as a recording medium 103a. The detection device 10 can read and write data from and to the recording medium 103a via the external I/F 103. Examples of the recording medium 103a include a flexible disk, a CD (Compact Disc), a DVD (Digital Versatile Disk), a SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

The communication I/F 104 is an interface for connecting the detection device 10 to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a storage device (storage device) such as a hard disk drive (HDD) or a solid state drive (SSD). The processor 108 is an arithmetic device such as a central processing unit (CPU) or a graphics processing unit (GPU).

The detection device 10 according to this embodiment has the hardware configuration shown in Fig. 1, and is therefore capable of implementing the meta-learning process and detection process described below. Note that the hardware configuration shown in Fig. 1 is merely an example, and the hardware configuration of the detection device 10 is not limited thereto. For example, the detection device 10 may have multiple auxiliary storage devices 107 and multiple processors 108, or may have various other hardware besides the hardware shown in the figure.

<Functional configuration of detection device 10>
An example of the functional configuration of the detection device 10 according to this embodiment is shown in Fig. 2. As shown in Fig. 2, the detection device 10 according to this embodiment has a meta-learning unit 201, a matching unit 202, a termination determination unit 203, and a detection unit 204. Each of these units is realized, for example, by a process in which one or more programs installed in the detection device 10 are executed by the processor 108. The detection device 10 according to this embodiment also has a storage unit 205. The storage unit 205 is realized, for example, by the auxiliary storage device 107. Note that the storage unit 205 may be realized, for example, by a storage device (such as a database server) connected to the detection device 10 via a communication network.

During meta-learning, the meta-learning unit 201 uses a learning data set D for meta-learning to learn a common model so as to improve the detection performance of out-of-distribution data.

The adaptation unit 202 trains a task-specific model to adapt to a certain training data set (the support set described later) during both meta-learning and detection.

The termination determination unit 203 determines whether or not a termination condition for the iterations for learning a common model is met during meta-learning.

At the time of detection, the detection unit 204 detects out-of-distribution data using the common model learned during meta-learning and the task-specific model for the target task t'.

The storage unit 205 stores at least the common model (and its parameters) to be learned, the task-specific model (and its parameters) for each task, and the learning data set D for meta-learning during meta-learning. Furthermore, the storage unit 205 stores at least the learned common model (and its parameters), the learning data set D _t' for the target task t', the task-specific model (and its parameters) for the target task t', and the feature amount (or a set thereof) to be determined as to whether or not the data is out-of-distribution data during detection. In addition to these, the storage unit 205 may store various data such as intermediate calculation results during meta-learning or detection.

<Meta-learning process>
The meta-learning process according to this embodiment will be described with reference to FIG.

First, the meta-learning unit 201 initializes the parameters Φ of the common model (hereinafter also referred to as the common model parameters Φ) (step S101). The common model parameters Φ may be initialized by any method (for example, random initialization, initialization according to a certain distribution, etc.).

Next, the meta-learning unit 201 randomly selects a task t ∈ {1, ..., T} (step S102).

Next, the meta-learning unit 201 randomly samples a support set S ⊂ D _t for the task t selected in step S102 (step S103). That is, the meta-learning unit 201 randomly samples some learning data (x _tn , y _tn ) from the learning data set D _t for the task t selected in step S102, and sets the set of these learning data as a support set S. Note that the number of samples of learning data may be determined randomly or may be a predetermined number.

Next, the meta-learning unit 201 randomly samples the in-distribution query set _QI of the task t selected in step S102 above (step S104). That is, the meta-learning unit 201 randomly samples some feature quantities x _tn from the set of feature quantities {x _tn |n=1, ..., N _t } included in the learning data set D _t of the task t selected in step S102 above, and sets the set of feature quantities as the in-distribution query set _QI . Note that the number of sampled feature quantities may be determined randomly or may be a predetermined number.

Next, the meta-learning unit 201 randomly samples the out-of-distribution query set Q _O of the task t selected in step S102 above (step S105). That is, the meta-learning unit 201 randomly samples some features from a set of features included in the learning data set D\D _t of a task other than the task t selected in step S102 above, and sets the set of features as the out-of-distribution query set Q _O. Note that the number of sampled features may be determined randomly or may be a predetermined number.

Next, the adaptation unit 202 estimates parameters of the task specific model for the task t selected in step S102 so as to adapt to the support set S obtained in step S103 (step S106). Hereinafter, the parameters of the task specific model are denoted as Θ. Note that since a density distribution model is assumed as the task specific model, the parameters can be estimated, for example, by maximum likelihood estimation, and a specific example will be described later.

Next, the meta-learning unit 201 calculates the out-of-distribution score for the query set Q=Q _I ∪Q _O using the task-specific model for the task t selected in step S102 above (step S107).

Here, the out-of-distribution score of feature x when the support set S and common model parameters Φ are given is defined as u(x|S;Φ). The out-of-distribution score u(x|S;Φ) takes a high value if the class of feature x is not included in the classes represented by the labels of each training data constituting the support set S, and takes a low value if the class of feature x is included. As will be described later, at the time of detection, the value of this out-of-distribution score is used to determine whether or not a given feature is out-of-distribution data. An example of such an out-of-distribution score u(x|S;Φ) is the negative log-likelihood in the latent space

を用いることができる。ここで、ｆは特徴量空間から潜在空間へ特徴量を写像するニューラルネットワーク、φはそのパラメータである。また、

Here, f is a neural network that maps features from the feature space to the latent space, and φ is its parameter.

はサポート集合Ｓに適合したときのタスク特化モデルのパラメータ（つまり、上記のステップＳ１０６で推定されたパラメータ）を表す。ニューラルネットワークｆのパラメータφは、共通モデルパラメータに含まれる（つまり、φ∈Φである。）。以下、明細書のテキスト中では、推定値を表すハット「＾」は記号の直前に記載する。例えば、サポート集合Ｓに適合したときのタスク特化モデルのパラメータを「＾Θ（Ｓ；Φ）」と表記する。

represents the parameters of the task-specific model when it matches the support set S (i.e., the parameters estimated in step S106 above). The parameter φ of the neural network f is included in the common model parameters (i.e., φ∈Φ). Hereinafter, in the text of the specification, a hat "^" indicating an estimated value is written immediately before the symbol. For example, the parameters of the task-specific model when it matches the support set S will be written as "^Θ(S;Φ)".

Therefore, for each x∈Q, the meta-learning unit 201 calculates, for example, an out-of-distribution score u(x|S;Φ) as shown in the above equation (1).

Here, when the negative log-likelihood is used as the out-of-distribution score u(x|S;Φ), any density distribution model can be used as the task-specific model. Therefore, for example, when a Gaussian mixture distribution is used as the task-specific model, in step S106 above, the adaptation unit 202 can calculate the parameters shown in the following formula (2) by maximum likelihood estimation.

ここで、Ｓ_ｋはサポート集合Ｓの中でクラスｋである特徴量の集合、＾γ_ｋ（Ｓ）はクラスｋの混合比、＾μ_ｋ（Ｓ）はクラスｋの特徴量の潜在空間における平均、＾Σ_ｋ（Ｓ）はクラスｋの特徴量の潜在空間における共分散、βはパラメータ、Ｉは単位行列である。これらの混合比＾γ_ｋ（Ｓ）、平均＾μ_ｋ（Ｓ）及び共分散＾Σ_ｋ（Ｓ）をまとめたものがパラメータ＾Θ（Ｓ；Φ）、つまり＾Θ（Ｓ；Φ）＝｛＾γ_ｋ（Ｓ），＾μ_ｋ（Ｓ），＾Σ_ｋ（Ｓ）｜ｋ＝１，・・・，Ｋ_ｔ｝である。なお、混合ガウス分布ではなく、例えば、カーネル密度推定やニューラルネットワークに基づく密度モデル等を用いることも可能である。

Here, S _k is a set of features of class k in the support set S, ^γ _k (S) is the mixture ratio of class k, ^μ _k (S) is the average of the features of class k in the latent space, ^Σ _k (S) is the covariance of the features of class k in the latent space, β is a parameter, and I is a unit matrix. The mixture ratio ^γ _k (S), the average ^μ _k (S), and the covariance ^Σ _k (S) are combined to form the parameter ^Θ(S;Φ), that is, ^Θ(S;Φ)={^γ _k (S), ^μ _k (S), ^Σ _k (S)|k=1,...,K _t }. Note that instead of the mixed Gaussian distribution, it is also possible to use, for example, a density model based on kernel density estimation or a neural network.

Next, the meta-learning unit 201 evaluates the out-of-distribution score calculated in step S107 above (i.e., the out-of-distribution score for the query set Q) (step S108).

Here, for example, the AUC (area under the ROC curve) can be used as an evaluation measure for the out-of-distribution score (in other words, a measure for evaluating the detection performance of out-of-distribution data using the out-of-distribution score). In this case, the AUC for the query set Q (more precisely, the AUC for the query set Q when the parameter ^Θ(S;Φ) estimated in the above step S106 and the common model parameter Φ are given) is

と計算することができる。ここで、Ｎ_Ｏは分布外クエリ集合Ｑ_Ｏの大きさ（要素数）、Ｎ_Ｉは分布内クエリ集合Ｑ_Ｉの大きさ（要素数）を表す。また、Ｉ（・）は指示関数であり、Ａが真のときＩ（Ａ）＝１、そうでないときＩ（Ａ）＝０を取る。なお、上記の式（３）において、指示関数の代わりに、指示関数をシグモイド関数で近似した微分可能な関数を用いてもよい。

Here, N _O represents the size (number of elements) of the out-of-distribution query set Q _O , and N _I represents the size (number of elements) of the in-distribution query set Q _I . In addition, I(·) is an indicator function, and I(A)=1 when A is true, and I(A)=0 when not. In addition, in the above formula (3), a differentiable function that approximates the indicator function with a sigmoid function may be used instead of the indicator function.

Next, the meta-learning unit 201 updates the common model parameter Φ so as to improve the detection performance of out-of-distribution data (step S109).

For example, the common model parameters Φ are learned so that the expected value of AUC for the query set Q is maximized. That is, the common model parameters Φ are learned using the following equation (4).

ここで、Ｅは期待値を表す。なお、共通モデルパラメータΦの更新は、例えば、既知の勾配法等により行うことができる。

Here, E represents an expectation value. Note that the common model parameter Φ can be updated by, for example, a known gradient method.

Next, the termination determination unit 203 determines whether a predetermined termination condition is satisfied (step S110). Examples of the termination condition include the number of repetitions of steps S102 to S109 being equal to or greater than a predetermined threshold, the amount of change in the objective function to be optimized (i.e., the expected value E in the above formula (4)) being equal to or less than a predetermined threshold, the amount of change in the model common parameter Φ being equal to or less than a predetermined threshold, etc.

If it is not determined in step S110 above that the predetermined termination condition is met, the process returns to step S102. As a result, steps S102 to S110 are repeatedly executed until it is determined that the termination condition is met. On the other hand, if it is determined in step S110 above that the predetermined termination condition is met, the meta-learning process is terminated. In this case, a trained common model has been obtained.

<Detection process>
The detection process according to this embodiment will be described with reference to FIG.

The adaptation unit 202 adapts the task-specific model of the target task t' to the support set S by using the training data set D _t' of the target task t' as the support set S (step S201). That is, the adaptation unit 202 estimates parameters ^Θ(S;Φ) of the task-specific model of the target task t' so as to adapt to the support set S, similar to step S106 in Fig. 3, where Φ is a parameter of the trained common model.

Next, the detection unit 204 calculates the out-of-distribution score u(x|S:Φ) of the feature x to be determined as being out-of-distribution data (step S202). That is, the detection unit 204 calculates the out-of-distribution score u(x|S:Φ), for example, by the above formula (1). Note that when a set of features {x _n |n=1, ..., N} to be determined as being out-of-distribution data is given, the out-of-distribution score u(x _n |S:Φ) may be calculated for each of n=1, ..., N.

Then, the detection unit 204 judges whether the corresponding feature is out-of-distribution data from the out-of-distribution score calculated in step S202 above (step S203). Specifically, for example, if the out-of-distribution score u(x|S:Φ) is equal to or greater than a predetermined threshold, the feature x is judged to be out-of-distribution data, and if not, it is judged not to be out-of-distribution data. As a result, the feature x whose out-of-distribution score u(x|S:Φ) is equal to or greater than a predetermined threshold is detected as out-of-distribution data.

<Evaluation>
The results of evaluating the detection device 10 according to this embodiment using handwritten characters are shown in Table 1 below.

上記の表１中において、提案手法が検出装置１０である。一方で、既存手法は既存のメタ学習法であるＰｒｏｔｏｔｙｐｉｃａｌ ｎｅｔｗｏｒｋとした。また、評価指標としてはＡＵＣを用いた。上記の表１に示されるように、本実施形態に係る検出装置１０は、既存手法と比べて、高いＡＵＣを達成していることがわかる。

In Table 1 above, the proposed method is the detection device 10. On the other hand, the existing method is a prototypical network, which is an existing meta-learning method. In addition, AUC was used as the evaluation index. As shown in Table 1 above, it can be seen that the detection device 10 according to the present embodiment achieves a higher AUC than the existing method.

<Summary>
As described above, the detection device 10 according to the present embodiment uses a training data set of a plurality of tasks including a target task to learn parameters of a common model common to these plurality of tasks through meta-learning. Then, the detection device 10 according to the present embodiment can detect out-of-distribution data of a target task from given features using the trained common model and a task-specific model adapted to the training data set of the target task. Moreover, even if the training data set of the target task contains only a small amount of training data, it can detect out-of-distribution data with higher accuracy than existing methods.

Although the above describes the case where out-of-distribution data for a target task is detected from given features, the detection device 10 according to this embodiment may execute various processes based on the detection results. For example, when the out-of-distribution data represents abnormal data, control processes such as stopping the abnormal device or switching to another device may be executed.

The present invention is not limited to the specifically disclosed embodiments above, and various modifications, variations, and combinations with known technologies are possible without departing from the scope of the claims.

10 Detection device 101 Input device 102 Display device 103 External I/F
103a Recording medium 104 Communication I/F
105 RAM
106 ROM
107 Auxiliary storage device 108 Processor 109 Bus 201 Meta-learning unit 202 Matching unit 203 End determination unit 204 Detection unit 205 Storage unit

Claims (7)

a meta-learning procedure for learning a common model, which is a machine learning model common to the plurality of tasks, using a first learning data set composed of learning data of a plurality of tasks including a target task that is a target task;
a determination step of determining whether or not given data is out-of-distribution data using the trained common model, a second training data set composed of training data for the target task, and a target task-specific model which is a machine learning model for the target task;
The computer executes
The determination procedure includes:
A detection method comprising: calculating an out-of-distribution score for the data; determining that the data is out-of-distribution data if the out-of-distribution score is equal to or greater than a predetermined threshold; and determining that the data is not out-of-distribution data if the out-of-distribution score is less than the threshold . The detection method includes:
an adaptation step of training a task-specific model, which is a machine learning model of the task, using training data of each task included in the plurality of tasks;
The meta-learning procedure comprises:
calculating the out-of-distribution score for each of the one or more first data and the one or more second data using one or more first data for a task selected from the plurality of tasks, one or more second data for a task other than the one task, and the trained task-specific model for the one task ;
The method of claim 1 , further comprising training the common model such that the out-of-distribution scores improve the performance of detecting the out-of-distribution data. The out-of-distribution score is the negative log-likelihood in the latent space,
The meta-learning procedure comprises:
Calculating the expected AUC as a measure of performance for detecting the out-of-distribution data;
The method of claim 2 , wherein the common model is trained to maximize an expected value of the AUC. the task specific model is a predetermined density distribution model;
The adaptation procedure comprises:
The method of claim 3 , further comprising training the task-specific model by estimating parameters of the task-specific model by maximum likelihood estimation. The adaptation procedure comprises:
training the target task-specific model using the second training data set;
The determination procedure includes:
Calculating the out-of-distribution score for a given data set using the trained target task-specific model;
The method of claim 2 , further comprising determining that the data is out-of-distribution data if the out- of -distribution score is equal to or greater than the threshold. a meta-learning unit configured to learn a common model, which is a machine learning model common to the plurality of tasks, using a first learning data set composed of learning data of a plurality of tasks including a target task that is a target task;
a determination unit configured to determine whether or not given data is out-of-distribution data by using the trained common model, a second training data set composed of training data for the target task, and a target task-specific model which is a machine learning model for the target task; and
having
The determination unit is
A detection device that calculates an out-of-distribution score for the data, and determines that the data is out-of-distribution data if the out-of-distribution score is greater than or equal to a predetermined threshold, and determines that the data is not out-of-distribution data if the out-of-distribution score is less than the threshold . A program for causing a computer to execute the detection method described in any one of claims 1 to 5.

Images