Model learning device, model learning method, and model learning program

Abstract

To provide a model learning device, a model learning method, and a model learning program that can improve the level of robustness assurance of learning models using FSL. [Solution] The model learning device 1 includes a calculation unit 11 that calculates the value of a new loss function by introducing at least one of the following terms into the loss function of the meta-learning model: a symmetric loss between a first dataset obtained by data augmenting the original training dataset and the original dataset, and a symmetric loss between a second dataset obtained by adding noise to the original dataset and the original dataset; an update unit 12 that updates the parameters of the meta-learning model to minimize the loss function; and an output unit 13 that repeats the processing of the calculation unit 11 and the update unit 12 a certain number of times and outputs the updated meta-learning model as a learning model with improved robustness.

Description

Technical Field

[0001] The present invention relates to a technique for improving the robustness guarantee level of a model in few-shot learning.

Background Art

[0002] Few-shot learning (FSL) is a machine learning method aimed at learning a general-purpose model from a small number of samples. In general machine learning, a large amount of learning data on the order of thousands or tens of thousands is required for model learning. However, depending on the task and target data, it may not be possible to collect sufficient samples. To solve this problem, in FSL, in the meta-learning phase, a meta-learning model is created that learns the general knowledge necessary to flexibly handle unknown tasks. By using this meta-learning model in the meta-test phase, it becomes possible to handle unknown tasks, such as image classification of new species of animals, with only a small amount of learning data. Furthermore, it has been reported in Non-Patent Document 1 that, in the meta-test phase, by performing fine-tuning with data augmentation on the meta-learning model, state-of-the-art accuracy can be achieved in, for example, image classification.

[0003] However, like general machine learning methods, FSL is also vulnerable to adversarial attacks and requires defense methods. Representative defense methods such as adversarial learning, detection of adversarial samples, and vulnerability testing that performs attacks in a simulated manner have a certain effect. However, since the types of attacks must be assumed in advance, it has been difficult to handle unknown attacks. In recent years, a defensive method called Certified Robustness (CR), which provides theoretical guarantees of the robustness of learning models, has been proposed in Non-Patent Document 2. With CR, when noise is added to certain data to change the data, it can be theoretically guaranteed that the model's prediction for that data will not change up to a certain distance from the original data. Because CR provides guarantees based on the distance of the transformed data, it does not depend on the algorithm of the attack method and can also guarantee the robustness of the model against unknown attacks. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] SX Hu et al., "Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference," Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2022. [Non-Patent Document 2] JM Cohen et al., "Certified Adversarial Robustness via Randomized Smoothing," International Conference on Machine Learning. PMLR, 2019. [Non-Patent Document 3] M. Pautov et al., "Smoothed Embeddings for Certified Few-Shot Learning," Advances in Neural Information Processing Systems 35 (2022): 24367-24379. [Non-Patent Document 4] J. Snell et al., "Prototypical Networks for Few-shot Learning," Advances in Neural Information Processing Systems 30 (2017). [Overview of the project] [Problems that the invention aims to solve]

[0005] In FSL, a method for achieving CR by adding Gaussian noise has been proposed in Non-Patent Document 3, but the current level of assurance is insufficient. For this reason, a method called Robust Training (RT), which involves adding noise during the meta-learning phase, has been proposed to improve the level of robustness assurance, but its effectiveness has been reported to be limited.

[0006] Thus, in FSL, there is no established effective method to sufficiently improve the level of robustness assurance of the model. Furthermore, since RT methods are applied in the meta-learning phase, applying fine-tuning with data augmentation to the meta-learning model in the meta-testing phase could potentially reduce the level of robustness assurance.

[0007] The present invention aims to provide a model learning device, a model learning method, and a model learning program that can improve the level of robustness assurance of learning models using FSL. [Means for solving the problem]

[0008] The model learning device according to the present invention comprises: a calculation unit that calculates the value of a new loss function by introducing at least one of the following terms into the loss function of a meta-learning model: a symmetric loss between a first dataset obtained by data augmenting the original training dataset and the original dataset, and a symmetric loss between a second dataset obtained by adding noise to the original dataset and the original dataset; an update unit that updates the parameters of the meta-learning model to minimize the loss function; and an output unit that repeats the processing of the calculation unit and the update unit a certain number of times and outputs the updated meta-learning model as a learning model with improved robustness.

[0009] The calculation unit may calculate a new loss function value by further introducing a comparison loss term between the original dataset, the first dataset, or the second dataset and the third dataset obtained by adding noise to the second dataset.

[0010] The aforementioned source data may be a support set used in the fine-tuning stage of the meta-testing phase in FSL.

[0011] The aforementioned source data may be a queryset used in the meta-learning phase of FSL.

[0012] The model learning method according to the present invention involves a computer calculating the value of a new loss function by introducing at least one of the following terms into the loss function of the meta-learning model: a calculation unit, the symmetric loss between a first dataset obtained by data augmenting the original training dataset and the original dataset, and a symmetric loss between a second dataset obtained by adding noise to the original dataset and the original dataset; an update unit, the parameters of the meta-learning model, to minimize the loss function; repeating the processing of the calculation unit and the update unit a certain number of times; and outputting the updated meta-learning model as a learning model with improved robustness by an output unit.

[0013] The model learning program according to the present invention is for causing a computer to function as the model learning device. [Effects of the Invention]

[0014] According to the present invention, the level of robustness assurance for learning models using FSL is improved. [Brief explanation of the drawing]

[0015] [Figure 1] This is a block diagram showing the functional configuration of a model learning device in an embodiment. [Figure 2] This figure illustrates an example of the algorithm for the model learning method in the embodiment. [Figure 3] This is a conceptual diagram illustrating the configuration of the loss function in the embodiment. [Modes for carrying out the invention]

[0016] An example of an embodiment of the present invention will be described below. The model learning method of this embodiment introduces a contrasting loss term into the loss function used during model training in order to improve the level of robustness assurance of the meta-learning model compared to conventional FSL. First, the conventional FSL procedure targeted by the model learning method of this embodiment is shown.

[0017] [FSL] The conventional FSL that is improved by this embodiment will be described. FSL has two phases: the meta-learning phase and the meta-testing phase. In the meta-learning phase, the model learning system prepares multiple data sets called "episodes" and uses them to train the model with general knowledge. Episodes are support sets.

number

number

[0018] Here, if we set the set composed of N types of classes included in a certain episode as C = {c1,…, c } s i ∈ C, y q i ∈ C, and they are the labels of the support image s i and the query image q i respectively. The support set S is used to learn the generalized representation of each class. The model learning device calculates the loss function in the query set Q using this representation and adjusts the model parameters to minimize the loss.

[0019] The number of images constituting the support set S is determined by the number of classes N to be classified and the number of images of each class K used for learning. The scenario at this time is called N-way-K-shot. For example, in the case of 5-way-5-shot with N = 5 and K = 5, the number of images constituting the support set S is N × K = 5 × 5 = 25. Since the query set Q is a dataset for parameter update, it is common that N is set to the same value as the support set S and K is set to a value larger than the value in S.

[0020] As a typical method of FSL, a method (ProtoNet) that performs learning based on the representative points of each class called prototypes is proposed in Non-Patent Document 4. The prototype is created based on the embedded representation of the images in the support set S. The loss function L protonet (Q) used for the learning of ProtoNet for the query set Q is expressed as follows with the number of images in Q being N q .

Equation

[0021] Here, h cThis is the support set S belonging to class c. c It is a prototype generated based on,

number

number

[0022] In the meta-testing phase, the model learning system evaluates the classification accuracy on data from classes not used in the meta-learning phase to determine how well the meta-learned model can be applied to classifying unknown classes.

[0023] [Data expansion] In FSL, it is assumed that the number of image samples available for training is small, so ensuring diversity in image data is important to improve the accuracy of tasks such as image classification. One method widely used to achieve this is data augmentation. Data augmentation allows for the generation of diverse representations of an image data, such as changing the color tone of an image, cropping a part of an image, or rotating an image.

[0024] [Fine tuning] In the field of machine learning, fine-tuning, which involves adjusting (retraining) the model's parameters, is sometimes performed to improve the model's performance on a particular task. While meta-learning models in FSL are trained to be flexibly applicable to unknown tasks, and therefore fine-tuning may not always be necessary, Non-Patent Document 1 reports that fine-tuning using augmented image data can further improve the model's classification performance.

[0025] [Functional Configuration] Figure 1 is a block diagram showing the functional configuration of the model learning device 1 in this embodiment. The model learning device 1 is an information processing device (computer) equipped with a control unit 10 and a storage unit 20, as well as various input / output interfaces.

[0026] The control unit 10 is the part that controls the entire model learning device 1, and realizes the functions described later in this embodiment by appropriately reading and executing various programs stored in the memory unit 20. The control unit 10 may be a CPU. The control unit 10 comprises a calculation unit 11, an update unit 12, and an output unit 13. These functional units enable the updating of the meta-learning model to improve the level of robustness assurance.

[0027] The memory unit 20 is a storage area for various programs and data necessary for the hardware group to function as a model learning device 1, and may be ROM, RAM, flash memory, or a hard disk drive (HDD).

[0028] The calculation unit 11 calculates the value of a new loss function for the meta-learning model in FSL by introducing at least one of the following terms: the symmetric loss between the first dataset obtained by augmenting the original training dataset and the original dataset, and the symmetric loss between the second dataset obtained by adding noise to the original dataset and the original dataset. Furthermore, the calculation unit 11 may calculate a new loss function value by further introducing a comparison loss term between the third dataset obtained by adding noise to the second dataset and the original dataset, the first dataset, or the second dataset.

[0029] Here, the source data may be either a support set used in the fine-tuning stage of the meta-testing phase in FSL, or a query set used in the meta-learning phase.

[0030] The update unit 12 updates the parameters of the meta-learning model to minimize the loss function. The output unit 13 repeats the processing of the calculation unit 11 and the update unit 12 a certain number of times and outputs the updated meta-learning model as a learning model with improved robustness.

[0031] [Proposed method] The model learning device 1 of this embodiment creates a model that is resistant to data augmentation or protective noise for CR by introducing a controlled loss into the loss function used in the fine-tuning stage during the meta-learning phase or meta-testing phase of the model. Examples of control losses include supervised control losses and unsupervised control losses, as follows:

[0032] • Supervised control loss A set of images (support set or query set) is defined as X={(x1,y1),…,(x N´ ,y N´ Let )}. Here, N' is the number of images in X. X is a set of two images created by performing some kind of processing on X, such as data augmentation or noise addition. 1 ={(x 1 1, y 1 1), ..., (x 1 N´ ,y 1 N´ )} and X 2 ={(x 2 1, y 2 1), ..., (x 2 N´ ,y 2 N´ The supervised control loss between )} is calculated as follows:

number

number

[0033] • Unsupervised control loss On the other hand, the unsupervised control loss is calculated as follows:

number

number

[0034] [Examples of implementations of the proposed method] As an example, we will show an implementation example where fine-tuning with data augmentation is performed using ProtoNet during the metatesting phase.

[0035] The set obtained by adding Gaussian noise or other noise for the expected CR to the support set S is S smLet =smooth(S). Also, the set obtained by applying data augmentation to the support set S is the pseudo-queryset Q. aug Let =aug(S). Furthermore, Q aug For this, the set obtained by adding noise for CR is Q. sm =smooth(Q aug )

[0036] For example, S and Q aug Based on this, a supervised control loss is calculated using S and S sm When an unsupervised controlled loss is calculated based on the above, the final loss function is expressed as follows:

number

[0037] Figure 2 is a diagram illustrating the algorithm of the model learning method in this embodiment. Here, we show the algorithm corresponding to the aforementioned implementation example, and the meta-learning model f θ And using the support set S as input, the model learning device 1 learns a learning model f with improved robustness. θ´ Outputs.

[0038] In step 1, the model learning device 1 initializes the model parameters. In step 2, the model learning device 1 initializes the index n of the iteration process. In steps 3 to 11, the model learning device 1 performs a predetermined number of repetitions N. iter While counting up n until (step 10), the parameter update process (steps 4-9) is executed.

[0039] In step 4, the model learning device 1 uses the embedding representation H of the support set S. S Obtain it. In step 5, the model learning device 1 applies data augmentation processing to the support set S and creates a pseudo-query set Q augObtain it. In step 6, the model learning device 1 applies noise to the support set S. sm Obtain it. In step 7, the model learning device 1 uses a noisy pseudo-queryset Q sm Obtain it.

[0040] In step 8, the model learning device 1 assigns two types of control losses along with weight coefficients to the loss function L RT Calculate the value. In step 9, the model learning device 1 uses the gradient method or similar to calculate the loss function L RT Update the model parameters to minimize [the specified value].

[0041] In step 12, the model learning device 1 updates the model f θ´ This is output as a meta-learning model with improved robustness.

[0042] Figure 3 is a conceptual diagram illustrating the configuration of the loss function in this embodiment. In the aforementioned implementation example, a data-enhanced pseudo-queryset Q is created in image space based on the support set S. aug And S and Q aug Dataset S with Gaussian noise added to each of the following: sm and Q sm and are generated. These datasets are each transformed into embedding representations by the learned model.

[0043] For example, in the metatesting phase, if fine tuning is performed using ProtoNet, the prototype calculated from S and Q are used in the embedding space. sm Loss function L based on comparison with protonet This is used. In this embodiment, L protonet In contrast, S and S sm The contrasting losses, and S and Q, are shown in comparison. aug The target loss, compared to the other, is assigned a predetermined weight.

[0044] Note that the set of datasets to which the target loss is applied is not limited to the example combinations shown. In other words, in this embodiment, the calculation of the control loss term is not limited to the example algorithm described above; it is possible to change the combination of datasets to be compared, add control loss terms, change the values ​​of the weight coefficients, etc. For example, S sm and Q sm The contrast loss, Q aug and Q sm The contrasting losses, S and Q, are shown in comparison. sm A contrast loss, such as one that compares the two, may be used in place of or in addition to the examples described above. The choice of which contrast loss to use may be appropriately determined depending on the target learning model, the task being used, and the dataset.

[0045] Furthermore, since the loss function used for fine-tuning is generally the same as the loss function used in the meta-learning phase, the model learning method of this embodiment is also applicable to the meta-learning phase. In the example described above, since fine-tuning in the meta-testing phase was assumed, only the dataset derived from the support set S was used, but in the meta-learning phase, it is also possible to use the query set Q to calculate the control loss.

[0046] According to this embodiment, the model learning device 1 introduced a counter loss to the loss function used during fine tuning in the meta-learning phase or meta-testing phase during the FSL learning process. As a result, the model learning device 1 was able to acquire data augmentation and resistance to defensive noise to ensure the robustness of the model being learned. As a result, the model learning device 1 was able to improve the level of robustness assurance of the learned model in FSL.

[0047] Furthermore, the model learning device 1 can also introduce a similar control loss during both the meta-learning phase and the fine-tuning phase. This ensures consistency in the loss function configuration throughout the series of learning procedures, which is expected to improve model performance.

[0048] Furthermore, this will help suppress, for example, hostile attacks on learning models, thereby contributing to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs): "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0049] Although embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above. Furthermore, the effects described in the embodiments described above are merely a list of the most preferred effects resulting from the present invention, and the effects of the present invention are not limited to those described in the embodiments.

[0050] The model learning method by the model learning device 1 is implemented by software. When implemented by software, the programs constituting this software are installed on an information processing device (computer). These programs may be distributed to users by being recorded on removable media such as a CD-ROM, or by being downloaded to the user's computer via a network. Furthermore, these programs may be provided to the user's computer as a web service via a network without being downloaded. [Explanation of Symbols]

[0051] 1. Model Learning Device 10 Control Unit 11 Calculation Section 12 Update section 13 Output section 20 Memory section

Claims

1. A calculation unit calculates the value of a new loss function for the loss function of a meta-learning model by introducing at least one of the following terms into the loss function: a symmetric loss between a first dataset obtained by augmenting the original training dataset and the original dataset, and a symmetric loss between a second dataset obtained by adding noise to the original dataset and the original dataset. An update unit updates the parameters of the meta-learning model to minimize the loss function, A model learning device comprising: an output unit that repeats the processing of the calculation unit and the update unit a certain number of times and outputs the updated meta-learning model as a learning model with improved robustness.

2. The model learning apparatus according to claim 1, wherein the calculation unit calculates a new loss function value obtained by adding noise to the second dataset to a third dataset and further introducing a comparison loss term between the original dataset, the first dataset, or the second dataset.

3. The model learning apparatus according to claim 1 or 2, wherein the source data is a support set used in the fine-tuning stage of the meta-testing phase in FSL.

4. The model learning apparatus according to claim 1 or 2, wherein the source data is a queryset used in the meta-learning phase of FSL.

5. Computers The calculation unit calculates a new loss function value for the meta-learning model's loss function by introducing at least one of the following terms: the symmetric loss between the first dataset obtained by data augmenting the original training dataset and the original dataset, and the symmetric loss between the second dataset obtained by adding noise to the original dataset and the original dataset. The update unit updates the parameters of the meta-learning model to minimize the loss function. A model learning method comprising repeating the processing of the calculation unit and the update unit a certain number of times, and outputting the updated meta-learning model as a learning model with improved robustness using an output unit.

6. A model learning program for causing a computer to function as a model learning device according to claim 1 or claim 2.

Images