Synthesis device

Abstract

This synthesis device synthesizes a plurality of tuned models. First, in relation to a task vector indicating the difference in parameter values caused by tuning each model, the synthesis device calculates, on the basis of the absolute value of a task vector and the magnitude of the deviation (for example, the magnitude of the dispersion) between task vectors of a target parameter of said task vector, an importance score for the task vector. Next, for each parameter, the synthesis device selects, as the code of the parameter, the code of a value obtained by weighting the code of each task vector of the parameter by the importance score of said task vector and then adding the weighted codes. Afterward, the synthesis device synthesizes and outputs, for each parameter, a task vector having the same code as the code selected as the code of the parameter.

Description

Synthesizer

[0001] The present invention relates to a synthesis apparatus for synthesizing multiple finely tuned models.

[0002] A method that integrates multiple models that have been fine-tuned for specific tasks using Transfer Learning to create a multi-task model has attracted attention. However, a problem with the above method is that performance degrades due to interference between the parameters of the integrated models. Therefore, TIES-MERGING has been proposed as a method to resolve the interference between the parameters of the integrated models and generate a high-performance multi-task model (see Non-Patent Literature 1). TIES-MERGING is a model synthesis method consisting of three steps: Trim, Elect Sign, and Disjoint Merge.

[0003] Prateek Yadav et al., TIES-MERGING: Resolving Interference When Merging Models, 37th Conference on Neural Information Processing Systems (NeurIPS 2023)

[0004] However, the TIES-MERGING method described above had the problem that the accuracy of the model after synthesis for each task was not necessarily high. Therefore, the present invention aims to solve the above problem and improve the accuracy of the synthesized model for each task.

[0005] To solve the aforementioned problems, the present invention is characterized by comprising: a score calculation unit that calculates a score of importance of a task vector for a parameter based on the absolute value of the task vector and the magnitude of the difference between the task vector of the task vector and each task vector of the target parameter, for each of the tuned models to be synthesized; a selection unit that selects the sign of the parameter by adding the signs of each task vector of the parameter, weighted by the score of importance of the task vector, and then selecting the sign of the value obtained from that addition as the sign of the parameter for each parameter; and a synthesis unit that synthesizes and outputs a task vector for each parameter that has the same sign as the sign selected as the sign of the parameter.

[0006] According to the present invention, the accuracy of each task performed by the synthesized model can be improved.

[0007] Figure 1 is a diagram illustrating the overview of TIES-MERGING, which is a prerequisite technology for the model synthesis apparatus. Figure 2 is a diagram showing an example of the configuration of the model synthesis apparatus. Figure 3 is a flowchart showing an example of the processing procedure performed by the model synthesis apparatus. Figure 4 is a diagram showing the results of an evaluation experiment of a model synthesized by the model synthesis apparatus. Figure 5 is a diagram showing the results of an evaluation experiment of a model synthesized by the model synthesis apparatus. Figure 6 is a diagram showing an example of a computer that runs the model synthesis program.

[0008] The following describes embodiments for carrying out the present invention with reference to the drawings. The present invention is not limited to these embodiments.

[0009] [Prerequisite Technology] First, using Figure 1, we will explain the outline of TIES-MERGING, which is the prerequisite technology for the synthesis apparatus (model synthesis apparatus) of this embodiment. As mentioned above, TIES-MERGING consists of three steps: (1) Trim, (2) Elect Sign, and (3) Disjoint Merge.

[0010] Here, an example will be described where a computer synthesizes Model 1, 2, and 3 that are fine-tuned for each task. For example, for each of Model 1, 2, and 3 (task), the computer creates a Task Vector τ t = θ t - θ int (a vector indicating the difference in the parameter θ due to fine-tuning). Then, the computer sets τ t to 0 by leaving the top k% with the largest absolute value among the Task Vectors (task vectors) τ t ( (1) Trim).

[0011] After that, the computer calculates the sum of the task vectors τ t for each parameter, and selects the sign of the sum as the sign of the corresponding parameter ( (2) Elect Sign). Next, the computer synthesizes the task vectors τ t for each parameter. At this time, the computer adopts and synthesizes those task vectors τ t whose signs are the same as the sign selected in (2) ( (3) Disjoint Merge).

[0012] The model synthesis device of this embodiment improves the above (2) Elect Sign to improve the execution accuracy of each task by the synthesized model.

[0013] [Overview] Next, the overview of the model synthesis device will be described. Hereinafter, the task vector of parameter i of task t is referred to as vector v t,i .

[0014] For example, for the task vector v t,i remaining after the above (1) Trim, the model synthesis device calculates the importance score S t,i using the deviation (e.g., variance) from the group of task vectors of parameter i and the absolute value (L1 norm, ||·||1) of the task vector v t,i .

[0015] For example, the model synthesis device calculates the importance score S t,i of the task vector v t,iThis calculates the following: In other words, task vectors with a large deviation from the task vector group for parameter i, and with a large absolute value, can be considered to be task vectors with parameters that are highly important to task t (have values ​​specific to task t). Therefore, the model synthesis device calculates a higher importance score for task vectors with large variance and large absolute values.

[0016]

[0017] Subsequently, the model synthesis device uses the above importance score S t,i The normalized importance score S' t,i The model synthesis device then calculates the task vector v for each parameter i. t,i The sign of the value corresponds to the normalized importance score 'S' mentioned above. t,i The weighted values ​​are added together for each task t, and the sign of the added value is selected as the sign of the parameter i. For example, the model synthesis system selects the sign of parameter i based on the following equation (2).

[0018]

[0019] The model synthesis device selects the sign of each parameter based on the sign of the weighted and added value obtained by assigning importance scores to the signs of each task, as described above. This allows for adjustment of the balance between tasks in the synthesis of a finely tuned model for each task. As a result, the model synthesis device can improve the execution accuracy of each task by the synthesized model (multitasking model).

[0020] [Configuration Example] Next, an example of the configuration of the model synthesis apparatus 10 will be described using Figure 2. The model synthesis apparatus 10 includes, for example, an input / output unit 11, a storage unit 12, and a control unit 13.

[0021] The input / output unit 11 is an interface that handles the input and output of various types of data. For example, the input / output unit 11 accepts input of the parameter values ​​of each model before fine tuning and the parameter values ​​of each model after fine tuning.

[0022] The memory unit 12 stores data, programs, etc., that are referenced when the control unit 13 performs various processes. The memory unit 12 is implemented using semiconductor memory elements such as RAM (Random Access Memory) or flash memory, or storage devices such as hard disks or optical discs. For example, the memory unit 12 stores task vectors, etc.

[0023] The control unit 13 is responsible for controlling the entire model synthesis apparatus 10. The functions of the control unit 13 are realized, for example, by the CPU (Central Processing Unit) executing a program stored in the memory unit 12.

[0024] The control unit 13 includes an information input unit 131, a vector creation unit 132, a deletion unit 133, a code selection unit 134, a synthesis unit 137, and an information output unit 138.

[0025] The information input unit 131 receives input of information for the model synthesis device 10 to perform various processes. For example, the information input unit 131 receives input of the parameter values ​​of each model before fine tuning and the parameter values ​​of each model after fine tuning.

[0026] Furthermore, the information input unit 131 accepts input values ​​(the values ​​of α and β in equation (1)) that indicate the ratio in which the absolute value and variance of the task vector are added when the code selection unit 134 calculates the aforementioned importance score.

[0027] The vector generation unit 132 uses the parameter values ​​of each model before fine tuning and the parameter values ​​of each model after fine tuning to create task vectors for each parameter of each model.

[0028] The deletion unit 133 deletes redundant task vectors from the task vectors created by the vector creation unit 132. For example, the deletion unit 133 keeps a predetermined percentage of task vectors with the largest absolute values ​​and deletes the remaining task vectors. In other words, the deletion unit 133 performs the (1) Trim described above.

[0029] The symbol selection unit 134 selects the symbol of each parameter using the task vector remaining after deletion by the deletion unit 133. This symbol selection unit 134 includes a score calculation unit 135 and a selection unit 136.

[0030] The score calculation unit 135 calculates the magnitude of the deviation (for example, the value of variance) from the task vector group of the parameter targeted by the task vector remaining after deletion by the deletion unit 133. Then, the score calculation unit 135 calculates the importance score S of the task vector using the absolute value of the task vector and the magnitude of the deviation from the task vector group of the parameter targeted by the task vector.

[0031] Note that the score calculation unit 135 calculates such that the value of the importance score S of the task vector becomes higher as the absolute value of the task vector is larger and the deviation from the task vector group of the parameter targeted by the task vector is larger.

[0032] For example, the score calculation unit 135 calculates the importance score S by adding the absolute value of the task vector and the magnitude of the variance with respect to the task vector group of the parameter targeted by the task vector.

[0033] For example, the score calculation unit 135 calculates the importance score S of the task vector of parameter i of task t according to the above formula (1). Note that the values of α and β in formula (1) are values that can be appropriately set within the range where α + β = 1 by the user of the model synthesis device 10 (for example, α = 0.7, β = 0.3, etc. are set). t,i For example, if you want to strongly reflect the above-mentioned magnitude of variance in the importance score S, you can increase the value of β. If you do not want to strongly reflect the above-mentioned magnitude of variance in the importance score S, you can decrease the value of β.

[0034] For example, importance score S t,i For example, if you want to strongly reflect the above-mentioned magnitude of variance in the importance score S, you can increase the value of β. If you do not want to strongly reflect the above-mentioned magnitude of variance in the importance score S, you can decrease the value of β. t,i For example, if you want to strongly reflect the above-mentioned magnitude of variance in the importance score S, you can increase the value of β. If you do not want to strongly reflect the above-mentioned magnitude of variance in the importance score S, you can decrease the value of β.

[0035] The score calculation unit 135 normalizes the importance score S calculated by formula (1). For example, the score calculation unit 135 normalizes the importance score S based on the following formula (3). t,i For example, the score calculation unit 135 normalizes the importance score S based on the following formula (3). t,iThe normalized importance score S' t,i Calculate the result. Note that ε in equation (3) is a small constant to prevent division by zero.

[0036]

[0037] The selection unit 136, for each parameter, weights the signs of each task vector of that parameter by the importance score of that task vector and adds them together, and selects the sign of the resulting value as the sign of that parameter. For example, the selection unit 136 uses the above-mentioned formula (2) to select the sign s of parameter i. i Select the following.

[0038] Furthermore, if the summed value (sign) of any parameter becomes 0, the selection unit 136 selects the sign of the median of the task vector for that parameter as the sign of that parameter (see equation (4)).

[0039]

[0040] As a result, the selection unit 136 can stably select a code even for parameters with strong interference.

[0041] The following is an example of sign selection for each parameter. Parameter 0: Original values: ['0.0006', '0.0001', '0.0001'] Original signs: [1.0, 1.0, 1.0] Importance scores: ['0.0021', '0.0004', '0.0001'] Weighted signs: ['0.0021', '0.0004', '0.0001'] Final sign: +

[0042] Parameter 1: Original values: ['0.0007', '-0.0006', '-0.0012'] Original signs: [1.0, -1.0, -1.0] Importance scores: ['0.0024', '0.0040', '0.0008'] Weighted signs: ['0.0024', '-0.0040', '-0.0008'] Final sign: -

[0043] The merging unit 137, for each parameter of the task vectors remaining after deletion, merges and outputs task vectors that have the same sign as the parameter selected by the sign selection unit 134. In other words, the merging unit 137 performs the (3) Disjoint Merge described above.

[0044] The information output unit 138 uses the task vector output by the synthesis unit 137 to determine and output the values ​​of each parameter of the synthesized model (multitask model).

[0045] This model synthesis device 10 makes it possible to synthesize multitask models with high execution accuracy for each task.

[0046] [Example of Processing Procedure] Next, an example of a processing procedure performed by the model synthesis device 10 will be explained using Figure 3. First, the information input unit 131 of the model synthesis device 10 receives input of information for the model synthesis device 10 to perform various processes (S1). For example, the information input unit 131 receives input of the parameter values ​​of the model to be synthesized before fine tuning and the parameter values ​​after fine tuning. It also receives input of values ​​such as α and β in the equation (1) described above.

[0047] After S1, the vector generation unit 132 creates task vectors for each parameter of each model using the parameter values ​​before and after fine-tuning of the model to be synthesized (S2).

[0048] After S2, the deletion unit 133 deletes redundant task vectors from the task vectors created by the vector creation unit 132 (S3). For example, the deletion unit 133 keeps a predetermined percentage of the task vectors with the largest absolute values ​​and deletes the remaining task vectors.

[0049] After S3, the code selection unit 134 calculates the importance score S of the task vectors remaining after deletion by the deletion unit 133, and selects the code of each parameter by weighting the codes of each task vector using the importance score S (S4).

[0050] After S4, the synthesis unit 137 synthesizes task vectors for each parameter remaining after deletion in S3, and outputs task vectors having the same sign as the parameter selected in S4 (S5: Synthesis of task vectors for each parameter).

[0051] After S5, the information output unit 138 uses the task vectors of each parameter output in S5 to determine and output the values ​​of each parameter in the synthesized model (multitask model) (S6).

[0052] By having the model synthesis device 10 perform the above processing, it is possible to synthesize a multi-task model with high execution accuracy for each task.

[0053] [Experimental Results] Next, an evaluation experiment was conducted to assess the execution accuracy of each task of the model synthesized by the model synthesis apparatus 10. The comparative example was TIES-MERGING. In this experiment, the sign of parameter i was selected based on the following equation (5) for TIES-MERGING.

[0054]

[0055] The datasets used in the experiment are as follows: • ErnestSDavis / winograd_wsc: A dataset related to text comprehension • allenai / quartz: A dataset related to scientific problems • google-research-datasets / paws: A dataset related to understanding text structure

[0056] The models used in the experiment were T5-base models fine-tuned on the three datasets listed above: • T5-FT-wsc • T5-FT-paws • T5-FT-quartz

[0057] The three models described above were synthesized using TIES-MERGING and ASM-MERGING (model synthesis method using model synthesis device 10). In ASM-MERGING, the coefficients used to calculate the importance score (α and β in equation (1)) were set to α = 0.7 and β = 0.3. The three datasets described above were used to evaluate the accuracy of each model.

[0058] The experimental results are shown in Figure 4. As shown in Figure 4, it was confirmed that the model synthesized using ASM-MERGING achieved higher accuracy than TIES-MERGING for all of the above datasets.

[0059] Figure 5 shows a comparison of the accuracy of ASM-MERGING when variance (the variance between tasks described in equation (1)) is used in calculating importance scores, and when normalization of importance scores (see equation (3)) is performed, and when normalization is not performed.

[0060] As shown in Figure 5, although importance scores were not normalized in ASM-MERGING, it was confirmed that using variance (inter-task variance as described in equation (1)) in the calculation of importance scores resulted in a slight improvement in accuracy for allenai / quartz compared to TIES-MERGING. Furthermore, it was confirmed that normalizing importance scores and using variance in the calculation of importance scores in ASM-MERGING resulted in a significant improvement in accuracy compared to TIES-MERGING.

[0061] [System Configuration, etc.] Furthermore, the components of each part shown in the diagram are functional concepts and do not necessarily need to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown in the diagram, and all or part of it can be configured by functionally or physically distributing and integrating in any unit according to various loads and usage conditions. In addition, all or any part of the processing functions performed by each device can be realized by a CPU and the program executed on that CPU, or by hardware using wired logic.

[0062] Furthermore, among the processes described in the embodiments described above, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0063] [Program] The model synthesis device 10 described above can be implemented by installing a program (model synthesis program) as packaged software or online software on a desired computer. For example, by having the above program run on an information processing device, the information processing device can be made to function as the model synthesis device 10. The information processing device referred to here includes mobile communication terminals such as smartphones, mobile phones and PHS (Personal Handyphone System), as well as terminals such as PDA (Personal Digital Assistant).

[0064] Figure 6 shows an example of a computer running a model synthesis program. Computer 1000 has, for example, memory 1010 and a CPU 1020. Computer 1000 also has 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.

[0065] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0066] The hard disk drive 1090 stores, for example, the OS 1091, application program 1092, program module 1093, and program data 1094. That is, the program that defines each process executed by the model synthesis device 10 is implemented as a program module 1093 in which executable code for a computer is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing a process similar to the functional configuration in the model synthesis device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0067] Furthermore, the data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes them.

[0068] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070.

[0069] 10 Model synthesis device (synthesis device) 11 Input / output unit 12 Storage unit 13 Control unit 131 Information input unit 132 Vector creation unit 133 Deletion unit 134 Code selection unit 135 Score calculation unit 136 Selection unit 137 Synthesis unit 138 Information output unit

Claims

1. A synthesis apparatus characterized by comprising: a score calculation unit that calculates a score of importance for a task vector in a parameter based on the absolute value of the task vector and the magnitude of the difference between the task vector of the task vector and each task vector of the target parameter, for each of the tuned models to be synthesized; a selection unit that selects the sign of the parameter by adding the signs of each task vector of the parameter, weighted by the score of importance of the task vector, for each parameter; and a synthesis unit that synthesizes and outputs a task vector for each parameter that has the same sign as the sign selected for that parameter.

2. The synthesis apparatus according to claim 1, characterized in that, if the sign of the value obtained by the addition for any parameter becomes 0, the selection unit selects the sign of the median task vector of the task vector group for the parameter as the sign of that parameter.

3. The synthesis apparatus according to claim 1, characterized in that the score calculation unit increases the score value of the importance of the task vector as the absolute value of the task vector increases and the deviation of the task vector from each task vector of the target parameter increases.

4. The synthesis apparatus according to claim 1, wherein the score calculation unit further normalizes the importance scores of the task vectors based on the calculation results of the importance scores of each task vector of the target parameters of the task vectors, and the selection unit weights the signs of the task vectors using the normalized importance scores of the task vectors.

Images