WO2024180744A1 - Combining device, combining method, and combining program - Google Patents

Abstract

In the present invention, an acquisition unit (15a) acquires a first model that has been trained using first training data and a second model that has been trained using second training data. An identification unit (15b) uses a weight with respect to input data of the first model and a weight of the second model to identify a weight of a combined model combining the first and second models, on the basis of the flatness of the gradient of a loss function for each weight.

Description

Synthesis apparatus, synthesis method, and synthesis program

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

　Conventionally, methods are known for generating a model that performs well with dataset A and dataset B when dataset A and dataset B are acquired independently. For example, continuous learning (see Non-Patent Document 1) is used to expand the scope of application of data, federated learning (see Non-Patent Document 2) is used to collect data in one place and learn without considering privacy concerns, and a method for synthesizing the same data (see Non-Patent Document 3) is known.

“Introduction to Continual Learning”, [online], ∞ WIki, [Retrieved February 6, 2023], Internet <URL: https://wiki.continualai.org/the-continualai-wiki/introduction-to-continual-learning> “Federated learning”, [online], WIKIPEDIA, [Retrieved February 6, 2023], Internet <URL: https://en.wikipedia.org/wiki/Federated_learning> Samuel K. Ainsworth, Jonathan Hayase, Siddhartha Srinivasa, “GIT RE-BASIN: MERGING MODELS MODULO PERMUTATION SYMMETRIES”, December 2022

However, with conventional technology, it can be difficult to generate a model that can be used with both datasets A and B from model A trained with dataset A and model B trained with dataset B. For example, with conventional technology, learning with dataset A and learning with dataset B cannot be performed independently, resulting in operational constraints.

The present invention has been made in consideration of the above, and aims to generate a model that can be used with both dataset A and dataset B from model A trained with dataset A and model B trained with dataset B.

In order to solve the above-mentioned problems and achieve the object, the synthesis device according to the present invention is characterized by having an acquisition unit that acquires a first model trained using first learning data and a second model trained using second learning data, and an identification unit that uses the weight of the first model for the input data and the weight of the second model to identify the weight of a synthesis model obtained by synthesizing the first model and the second model based on the flatness of the gradient of the loss function of each weight.

According to the present invention, it is possible to generate a model that can be used with both datasets A and B from model A trained with dataset A and model B trained with dataset B.

FIG. 1 is a schematic diagram illustrating the schematic configuration of a synthesis apparatus. FIG. 2 is a diagram for explaining the synthesis process. FIG. 3 is a flowchart showing the synthesis process procedure. FIG. 4 is a diagram for explaining the embodiment. FIG. 5 is a diagram for explaining the embodiment. FIG. 6 is a diagram illustrating a computer that executes a synthesis program.

Below, one embodiment of the present invention will be described in detail with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, in the drawings, the same parts are denoted by the same reference numerals.

[Outline of synthesis apparatus]
The synthesis device of this embodiment generates a synthetic model that can be used for both datasets A and B using model A trained on dataset A and model B trained on dataset B.

First, define a composite dataset D of dataset A and dataset B as follows:

In addition, in supervised learning, where l is a loss function and θ is the weight of the model for the input data, the weight θ _A of model A, the weight θ _B of model B, and the weight θ _A+B of the synthetic model trained with the synthetic data set are expressed by the following equations (2), (3), and (4), respectively.

If we can identify the operation ★ that satisfies the following formula (5), it will be possible to learn dataset A and dataset B separately, and then combine the learned weights to generate a composite model that can be used with dataset A + B.

Here, in Deep Learning, there is symmetry in the manipulation of the network structure. Therefore, the synthesis device uses the symmetry in the rearrangement of weights to perform permutation, which rearranges the weights so as not to change the model output.

In that case, the following equation (6) holds.

By replacing the above formula (6) with the rule shown in the following formula (7), the following formula (8) is realized.

Then, the synthesis device generates a synthetic model by determining the average of the weight of model A and the rearranged weight of model B as the weight of the synthetic model, as described below.

[Configuration of synthesis device]
Fig. 1 is a schematic diagram illustrating the schematic configuration of a synthesis device. Fig. 2 is a diagram for explaining synthesis processing. First, as illustrated in Fig. 1, a synthesis device 10 is realized by a general-purpose computer such as a personal computer, and includes an input unit 11, an output unit 12, a communication control unit 13, a storage unit 14, and a control unit 15.

The input unit 11 is realized using input devices such as a keyboard and a mouse, and inputs various instruction information such as a command to start processing to the control unit 15 in response to input operations by an operator. The output unit 12 is realized by a display device such as a liquid crystal display, a printing device such as a printer, etc.

The communication control unit 13 is realized by a NIC (Network Interface Card) or the like, and controls communication between the control unit 15 and an external device such as a server via a network. For example, the communication control unit 13 controls communication between the control unit 15 and a management device or the like that manages various information such as the datasets to be subjected to the synthesis process described below and the trained models.

The storage unit 14 is realized by a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 14 stores in advance the processing program that operates the synthesis device 10 and data used during execution of the processing program, or stores it temporarily each time processing is performed. The storage unit 14 may be configured to communicate with the control unit 15 via the communication control unit 13.

The control unit 15 is realized using a CPU (Central Processing Unit), NP (Network Processor), FPGA (Field Programmable Gate Array), etc., and executes a processing program stored in memory. As a result, the control unit 15 functions as an acquisition unit 15a and an identification unit 15b, as exemplified in FIG. 1, and executes synthesis determination processing. Note that each of these functional units may be implemented in different hardware. The control unit 15 may also include other functional units.

The acquisition unit 15a acquires a first model trained using the first learning data and a second model trained using the second learning data. For example, the acquisition unit 15a acquires model A, which is trained using data set A used in the synthesis process described below and is currently in operation, and model B, which is trained using data set B collected thereafter.

The acquisition unit 15a acquires these pieces of data via the input unit 11, or from a management device that manages various pieces of information via the communication control unit 13. The acquisition unit 15a may also store the acquired communication data in the storage unit 14. The acquisition unit 15a may also transfer this information to the identification unit 15b described below without storing it in the storage unit 14.

The identification unit 15b uses the weight of the first model for the input data and the weight of the second model to identify the weight of the combined model that combines the first model and the second model based on the flatness of the gradient of the loss function of each weight.

Specifically, the identification unit 15b rearranges the weights without changing the output of the second model, and identifies the weight of the composite model by averaging the rearranged weights and the weights of the first model.

For example, a method called weight matching is known that can perform high-speed sorting without using data. Weight matching optimizes the following equation (9).

The determination unit 15b therefore performs weight matching taking into account the flatness of the landscape of the weight loss function, i.e., the flatness of the gradient. For example, the determination unit 15b performs optimization by adding a term that represents the flatness of the gradient of the weight loss function, as shown in the following formula (10).

Specifically, the identification unit 15b searches for a permutation operator P that maximizes the following equation (11). Here, β is a constant that balances with flat.

Here, the composite dataset of two different datasets A and B can be expressed by the following equation (12). For example, when s = 0, it represents only dataset A, and when s = 1, it represents only dataset B.

Furthermore, the optimal solution for the weight of model A trained on data set A is expressed by the following formula (13), and the optimal solution for the weight of model B trained on data set B is expressed by the following formula (14).

Furthermore, for the synthetic data set, the following equation (15) holds:

If we simplify the loss functions of each term in the above formula (15) and write them as h, f, and g, respectively, we can express it as the following formula (16).

Since the expected value is a linear calculation, the determination unit 15b searches for the minimum value of h, which is defined as the sum of f and g.

Here, in conventional weight matching, a search is performed to find the optimal solution between f and g as close as possible. However, since the model output does not change, the magnitude of the loss itself does not change. In this case, the flatness of g is not taken into account.

Then, when the determination unit 15b brings the solutions of f and g closer together, it selects the flattest solution of g among the solutions of g with different flatness, as shown in FIG. 2. In FIG. 2, an intuitive image of a movable g is shown by a dashed line. Note that f is assumed to be flat because it is the result of a search using SGD (stochastic gradient descent).

In this way, the identification unit 15b generates a composite model by identifying the weight of the composite model based on the flatness of the gradient of the loss function of the weight.

In addition, when calculating the loss function, the identification unit 15b may use proxy data generated by data condensation. For example, the identification unit 15b generates proxy data using a data condensation technique called gradient matching. This makes it easier to identify the composite model.

The synthesis process of this embodiment can be applied to, for example, general AI. In particular, it has a high affinity with Deep Learning's specialties of image recognition, natural language processing, voice recognition, etc., and can be used in face recognition systems, etc. For example, if company A merges with company B after generating face recognition model A for employees of company A, it is possible to generate a synthetic model that can be used by both companies A and B so that employees of company B can also use the face recognition system.

[Composition Processing]
Next, a compositing process performed by the compositing device 10 according to the present embodiment will be described with reference to Fig. 3. The compositing process according to the present embodiment includes a detection process and a search process. The flowchart in Fig. 3 starts, for example, when an operation input is made to instruct the start of the compositing process.

First, the acquisition unit 15a acquires model A trained on dataset A and model B trained on dataset B (step S1). For example, the acquisition unit 15a acquires model A trained on dataset A and in operation. The acquisition unit 15a also acquires model B trained on dataset B acquired thereafter.

Next, the identification unit 15b uses the weight of model A and the weight of model B to identify the weight of a composite model obtained by combining model A and model B based on the flatness of the gradient of the loss function of each weight (step S2).

Specifically, the identification unit 15b rearranges the weights without changing the output of model B, and identifies the weight of the composite model by averaging the rearranged weights and the weights of model A. In this way, the composite model is generated.

At that time, the identification unit 15b generates proxy data using a data condensation technique such as gradient matching, and uses the proxy data when calculating the loss function.

The identification unit 15b also outputs the generated composite model to the operation device, for example, via the output unit 12. This completes the series of composite processes.

[effect]
As described above, the acquiring unit 15a acquires the first model trained using the first learning data and the second model trained using the second learning data. The identifying unit 15b identifies the weight of a composite model obtained by combining the first model and the second model, using the weight of the first model for the input data and the weight of the second model, based on the flatness of the gradient of the loss function of each weight.

Specifically, the identification unit 15b rearranges the weights without changing the output of the second model, and identifies the weight of the composite model by averaging the rearranged weights and the weights of the first model.

This makes it possible to generate a model that can be used with both datasets A and B from model A trained on dataset A and model B trained on dataset B.

In addition, when the identification unit 15d calculates the loss function, it uses the proxy data generated by data condensation. This relaxes the operating conditions and makes it easier to identify the composite model.

[Example]
4 and 5 are diagrams for explaining the examples. In this example, a synthetic model was generated by the synthesis process of the above embodiment using model A trained with MNIST and model B trained with FashionMNIST. At that time, the model was set to MLP, MNIST and FashionMNIST were used as datasets, and source code created with reference to "https://github.com/samuela/git-re-basin" was used. Then, the accuracy and loss of the synthetic model using a synthetic dataset combining both MNIST and FashionMNIST were evaluated.

The vertical axis in Figures 4 and 5 shows the synthesis ratio λ between model A trained by MNIST and model B trained by the Fashion model MNIST, and the vertical axis shows the accuracy of the synthesized model.

Here, model A shows an accuracy of nearly 100% for MNIST, but an accuracy of nearly 0% for FashionMNIST. Therefore, in Figures 4 and 5, model A has an accuracy of about 50% for the synthetic dataset. Also, model B shows an accuracy of nearly 100% for FashionMNIST, but an accuracy of nearly 0% for MNIST. Therefore, in Figures 4 and 5, model B also has an accuracy of about 50% for the synthetic dataset.

First, Figure 4 shows two examples of cases where flatness is not taken into account: no permutation shown by the dashed line, and permutation shown by the solid line. Also, examples are shown for each case during learning (Train, thick line) and operation (Test).

In addition, Figure 5 shows two examples of cases similar to those in Figure 4 when flatness is taken into account.

In both Figures 4 and 5, it was confirmed that the accuracy of the synthetic model was higher when permutation was performed (shown by the solid line) than when permutation was not performed (shown by the dashed line).

It was also confirmed that the accuracy of the synthetic model was higher when the flatness shown in Figure 5 was taken into account than when the flatness shown in Figure 4 was not taken into account.

[program]
A program in which the process executed by the synthesis device 10 according to the above embodiment is written in a language executable by a computer can also be created. As an embodiment, the synthesis device 10 can be implemented by installing a synthesis program that executes the above synthesis process as package software or online software on a desired computer. For example, the above synthesis program can be executed by an information processing device, so that the information processing device can function as the synthesis device 10. In addition, the information processing device also includes mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone System), as well as slate terminals such as PDA (Personal Digital Assistant). The function of the synthesis device 10 may also be implemented on a cloud server.

FIG. 6 is a diagram showing an example of a computer that executes a synthesis program. The computer 1000 has, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012. The ROM 1011 stores a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to a hard disk drive 1031. The disk drive interface 1040 is connected to a disk drive 1041. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1041. The serial port interface 1050 is connected to a mouse 1051 and a keyboard 1052, for example. The video adapter 1060 is connected to a display 1061, for example.

Here, the hard disk drive 1031 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. Each piece of information described in the above embodiment is stored, for example, in the hard disk drive 1031 or memory 1010.

The synthesis program is stored in the hard disk drive 1031, for example, as a program module 1093 in which instructions to be executed by the computer 1000 are written. Specifically, the program module 1093 in which each process executed by the synthesis device 10 described in the above embodiment is written is stored in the hard disk drive 1031.

In addition, data used for information processing by the synthesis program is stored as program data 1094, for example, in the hard disk drive 1031. Then, the CPU 1020 reads the program module 1093 and program data 1094 stored in the hard disk drive 1031 into the RAM 1012 as necessary, and executes each of the above-mentioned procedures.

The program module 1093 and program data 1094 related to the synthesis program are not limited to being stored in the hard disk drive 1031, but may be stored in a removable storage medium, for example, and read by the CPU 1020 via the disk drive 1041 or the like. Alternatively, the program module 1093 and program data 1094 related to the synthesis program may be stored in another computer connected via a network, such as a LAN (Local Area Network) or a WAN (Wide Area Network), and read by the CPU 1020 via the network interface 1070.

The above describes an embodiment of the invention made by the inventor, but the present invention is not limited to the descriptions and drawings that form part of the disclosure of the present invention according to this embodiment. In other words, other embodiments, examples, operational techniques, etc. made by those skilled in the art based on this embodiment are all included in the scope of the present invention.

REFERENCE SIGNS LIST 10 Synthesis device 11 Input unit 12 Output unit 13 Communication control unit 14 Storage unit 15 Control unit 15a Acquisition unit 15b Identification unit

Claims (5)

an acquisition unit that acquires a first model trained using the first learning data and a second model trained using the second learning data;
A determination unit that determines a weight of a composite model obtained by combining the first model and the second model based on a flatness of a gradient of a loss function of each weight by using a weight for the input data of the first model and a weight of the second model;
A synthesis apparatus comprising: The synthesis device according to claim 1, characterized in that the identification unit identifies the weight of the synthesis model by rearranging the weights without changing the output of the second model and averaging the rearranged weights and the weights of the first model. The synthesis device according to claim 1, characterized in that the determination unit uses proxy data generated by data condensation when calculating the loss function. A synthesis method performed by a synthesis device, comprising:
An acquisition step of acquiring a first model trained using the first training data and a second model trained using the second training data;
A step of specifying a weight of a composite model obtained by combining the first model and the second model based on the flatness of the gradient of a loss function of each weight using the weight of the first model for the input data and the weight of the second model;
A synthesis method comprising: An acquisition step of acquiring a first model trained using the first training data and a second model trained using the second training data;
A step of specifying a weight of a composite model obtained by combining the first model and the second model based on a flatness of a gradient of a loss function of each weight using a weight for the input data of the first model and a weight of the second model;
A synthesis program for causing a computer to execute the above.

Images