Information processing system, information processing method, and program

Abstract

This information processing system comprises a data discriminator and a learning parameter control unit. The data discriminator determines a domain of learning data input for fine-tuning. The learning parameter control unit varies model parameters to be fine-tuned according to the determined domain of the learning data.

Description

Information Processing System, Information Processing Method, and Program

[0001] The present invention relates to an information processing system, an information processing method, and a program.

[0002] In AI model development, the construction cost of the teacher dataset required for learning, such as data collection and annotation work, is an issue. Also, since the range observable with real data is limited, there is an issue from the perspective of comprehensiveness of learning data. Therefore, there is an increasing demand to develop AI models using synthetic data generated by a simulator or the like.

[0003] CVPR24, “Stronger, Fewer, & Superior: Harnessing Vision Foundation Models for Domain Generalized Semantic Segmentation”, Zhixiang Wei, Lin Chen1, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huai'an Chen, Jinjin Zheng, University of Science and Technology of China, Shanghai AI Laboratory.

[0004] When using synthetic data generated by a simulator or the like for an AI learning model, the performance generally deteriorates compared to a model learned with real data due to the quality gap and concept gap with real data. As a technique for eliminating the domain gap that causes such performance degradation, a learning technique called domain adaptation is known. However, the domain gap between synthetic data and real data is generally large, and it is difficult to overcome with such conventional domain adaptation techniques.

[0005] One technique for training models on synthetic data is a learning method called Rein. This method utilizes a pre-trained Vision Foundation Model (VFM) as a backbone, and aims to overcome domain gaps by training only the token-based, low-rank parameter portion using synthetic data and a small amount of real-world data. However, in actual use cases, it is difficult to secure a specific VFM with high generalization performance in advance. Therefore, there is a need for methods to build high-performing learning models when high-quality training data is scarce.

[0006] Therefore, this disclosure proposes an information processing system, information processing method, and program that can construct a high-performance learning model using training data from various domains.

[0007] According to this disclosure, an information processing system is provided, comprising: a data discriminator for determining the domain of training data input for fine tuning; and a training parameter control unit for differentiating the model parameters subject to fine tuning according to the determined domain of the training data. Furthermore, according to this disclosure, an information processing method is provided in which the information processing of the information processing device is performed by a computer, and a program is provided for causing the computer to implement the information processing of the information processing device.

[0008] This is a diagram illustrating a fine-tuning method using LoRA. This is a diagram showing an example of the configuration of the learning system of this disclosure. This is a diagram showing variations in fine-tuning depending on the domain of the training data. This is a diagram showing variations in fine-tuning depending on the domain of the training data. This is a diagram showing variations in fine-tuning depending on the domain of the training data. This is a diagram showing variations in fine-tuning depending on the domain of the training data. This is a diagram showing an example of the hardware configuration of the learning system.

[0009] Embodiments of the present disclosure will be described in detail below with reference to the drawings. In each of the following embodiments, the same parts will be denoted by the same reference numerals, and redundant descriptions will be omitted.

[0010] The explanation will proceed in the following order: [1. Fine-tuning using LoRA and its challenges] [2. Configuration of the learning system in this disclosure] [3. Variations of fine-tuning according to the domain of the learning data] [4. Example hardware configuration]

[0011] [1. Challenges of Fine Tuning Using LoRA] Figure 1 is a diagram illustrating a fine tuning method using Low-Rank Adaptation (LoRA).

[0012] LoRA refers to a technique that reduces computational cost and memory usage by inserting low-rank matrices (low-rank parts) into each layer of a model and training only those matrices. LoRA is used to efficiently fine-tune large-scale language models and image generation models.

[0013] Figure 1 shows an example of applying LoRA to a learning method called Rein. In the example in Figure 1, a pre-trained Vision Foundation Model (VFM) is used as the backbone to train the model on synthetic data, and the domain gap is overcome by training only the token-based low-rank parameter portion on synthetic data and a small amount of real data.

[0014] In Rein, low-rank tokens are inserted into each layer of a pre-trained VFM, freezing all of the VFM's parameters. By focusing on higher-order information with low-rank tokens and excluding the VFM from the training target, it becomes possible to utilize the VFM's high generalization performance, thereby bridging the gap between real-world and synthetic data. However, in actual use cases, it is difficult to secure a specific VFM with high generalization performance in advance, making this method, which focuses on utilizing VFMs, difficult to implement.

[0015] Therefore, in this disclosure, the learning process is performed on a composite domain consisting of real data and synthetic data, while switching the model's learning target. The configuration of the learning system necessary to realize this method is described below.

[0016] [2. Configuration of the Learning System of the Disclosure] Figure 2 is a diagram showing an example of the configuration of the learning system 1 of the Disclosure.

[0017] The learning system 1 is an information processing system for learning processes to fine-tune the model parameters of the learning model 40. The learning system 1 of this disclosure executes the learning process while dynamically switching the model parameters to be learned within the learning model 40 according to the domain of the learning data. This reduces the domain gap between learning data with different characteristics and enables the learning of an AI model using synthetic data.

[0018] For example, the learning system 1 includes a data storage unit 10, a data discriminator 20, a learning parameter control unit 30, a learning model 40, and a loss calculation unit 50. The data storage unit 10 stores multiple types of learning data from different domains. The learning model 40 performs learning (fine-tuning) by switching the model parameters to be learned according to the domain of the input learning data.

[0019] In the example in Figure 2, real-world data and synthetic data are used as training data. Real-world data refers to data from the real world acquired by sensors. Synthetic data refers to data artificially generated by a computer program. Examples of real-world data include images of subjects taken with a real camera. Examples of synthetic data include computer graphics (CG) images generated by simulators, etc.

[0020] Real-world data includes both annotated and unannotated data. In Figure 2, to distinguish between the two, the former is labeled "Real Data +" and the latter "Real Data." Synthetic data is annotated. The annotations on synthetic data are generated based on parameters used during the simulation.

[0021] For example, the data storage unit 10 includes a first real data storage unit 11, a composite data storage unit 12, and a second real data storage unit 13. The first real data storage unit 11 stores a dataset of annotated real data (annotated real dataset). The composite data storage unit 12 stores a dataset of composite data generated by a simulator or the like. The second real data storage unit 13 stores a dataset of real data that has not been annotated (unannotated real dataset).

[0022] The data storage unit 10 outputs multiple datasets from different domains as batch processing data. In the example in Figure 2, each input batch consists of three types of datasets: annotated real datasets, synthetic datasets, and unannotated real datasets.

[0023] The data discriminator 20 identifies the domain of the training data input for fine-tuning. Possible methods of identification include, for example, the data discriminator 20 reading identifiers that have been pre-embedded in the training data, or classifying the training data using a separately defined classification model.

[0024] In the example in Figure 2, three types of training data domains are shown: "real data +", "synthetic data", and "real data". The data discriminator 20 uses the method described above to determine the domain of each training data. The training parameter control unit 30 adjusts the model parameters of the training model 40 that are subject to fine-tuning according to the determined domain of the training data.

[0025] With this configuration, appropriate model parameters are selected for fine-tuning depending on the domain of the training data. Therefore, even when high-quality training data is scarce, a high-performance training model 40 can be constructed by effectively utilizing training data from various domains.

[0026] The learning model 40 includes a feature extraction unit 41 and a low-rank unit 42. The learning model 40 is configured as a neural network in which a low-rank matrix (low-rank unit 42) is inserted into an intermediate layer (feature extraction unit 41) that performs feature extraction. The learning model 40 learns the coefficients (model parameters) of the neural network based on the training data. The loss calculation unit 50 calculates the difference between the output of the learning model 40 and the annotation as the loss value.

[0027] For example, in fine-tuning using real data as training data, the learning parameter control unit 30 determines the model parameters representing the feature extraction unit 41 as the target of fine-tuning. With this configuration, the feature extraction unit 41, which is the deep part of the learning model 40, is trained based on real data with high-resolution features from the real world. As a result, the accuracy of feature extraction is improved.

[0028] The learning method differs depending on whether annotated real data is used for training (annotated real data) or unannotated real data is used.

[0029] For example, in fine tuning using annotated real data as training data, the learning parameter control unit 30 determines that both the model parameters representing the feature extraction unit 41 and the model parameters representing the low-rank unit 42 are targets for fine tuning. The loss calculation unit 50 acquires the annotations as Ground Truth and calculates the difference between the output of the learning model 40 and the annotations as the loss value. The learning model 40 fine-tunes its model parameters to reduce the loss value.

[0030] In this configuration, the entire model is trained based on annotated real-world data. Annotated real-world data contains high-resolution features from the real world, and the annotations are accurate. Therefore, training the entire model with such data simultaneously reduces bias in synthetic data and improves the accuracy of feature extraction.

[0031] When using synthetic data as training data, the range of model parameters to be trained is limited in order to avoid the influence of biases specific to synthetic data. For example, in fine tuning using synthetic data as training data, the training parameter control unit 30 determines that the model parameters representing the low-rank section 42 are the target of fine tuning, and excludes the model parameters representing the feature extraction section 41 from the target of fine tuning.

[0032] The Loss calculation unit 50 acquires annotations as Ground Truth and calculates the difference between the output of the learning model 40 and the annotations as the loss value. The learning model 40 fine-tunes its model parameters to minimize the loss value. Since synthetic data contains many biases, training only the superficial, low-rank part 42 of the learning model 40, rather than the deep part, can reduce the degradation of model performance caused by biases specific to synthetic data.

[0033] When using unannotated real data as training data, annotations cannot be directly obtained from the data storage unit 10. Therefore, pseudo-annotations generated by the training model 40 are used as Ground Truth. Generally, pseudo-annotations are less accurate than annotations assigned as default values, and this inaccuracy can act as noise, potentially affecting the training results. For this reason, the range of model parameters to be trained is limited.

[0034] For example, in fine tuning using unannotated real data as training data, the learning parameter control unit 30 excludes model parameters indicating the low-rank portion 42 from the fine tuning process. This configuration suppresses the decrease in accuracy of the low-rank portion 42 caused by using inaccurate pseudo-annotations.

[0035] The generation of pseudo-annotations and the calculation of losses can be performed as follows. For example, the Loss calculation unit 50 obtains a training model by performing exponential moving average processing on the learning model 40. The Loss calculation unit 50 inputs real data into the training model to obtain pseudo-annotations. The Loss calculation unit 50 calculates the difference between the output of the learning model 40 obtained by inputting real data into the learning model 40 and the pseudo-annotations as the loss value.

[0036] In this method, the teacher model is defined by the exponential moving average (EMA) model of the current learning model 40. The teacher model is updated according to equation (1) below. In equation (1), "θ" ｔ " indicates the model parameters of the teacher model. ｓ " indicates the model parameters of the current learning model 40. "α" indicates the movement coefficient. The movement coefficient α is defined as a value in the range of 0 to 1. θ ｔ =αθ ｔ + (1 - α)θ ｓ ... (1)

[0037] In a teacher model, it is expected that robust pseudo-annotations will be added to address noise and outliers by averaging over the training iterations. Since pseudo-annotations have low accuracy, training only the feature extraction unit 41 in the deep part of the model can reduce the noise of pseudo-annotations while improving the accuracy of feature extraction.

[0038] Depending on the accuracy of the pseudo-annotation, it is possible to train the entire model in the same way as with annotated real data. Furthermore, the update of the training model represented by equation (1) can target all model parameters of the feature extraction unit 41 and the low-rank unit 42, or it can target only the model parameters of the low-rank unit 42.

[0039] [3. Variations in fine-tuning according to the domain of the training data] Figures 3 to 6 show variations in fine-tuning according to the domain of the training data.

[0040] When using a plurality of different synthetic data for learning, the configuration can be such that the learning process is further branched according to its characteristics. For example, there is a difference in the amount of information contained between synthetic data generated using a game engine-based physical simulator and synthetic data reconstructed based on real data like NeRF. Generally, the former synthetic data has a large gap from real data, while the latter tends to have a small gap.

[0041] In such a case, for the former input with a large gap, only the low-rank part 42 is learned (see FIG. 3), and for the latter input with a small gap, a modified example can be considered where not only the low-rank part 42 but also a part of the feature extraction unit 41 is learned (see FIG. 4). Alternatively, for the former input with a large gap, only a part of the low-rank part 42 is set as the learning target (see FIG. 5), and for the latter input with a small gap, the entire low-rank part 42 is set as the learning target (see FIG. 6). In FIGS. 3 to 6, the areas of the model parameters to be learned are schematically shown by hatching.

[0042] Thus, by dynamically switching the depth and the number of parameters to be learned according to the characteristics of the synthetic data, it is possible to reduce the performance degradation of the learning model 40 even for the gaps of synthetic data with different degrees.

[0043] Note that the method of the present disclosure does not necessarily require both types of real data (annotated real data and non-annotated real data) distinguished by the presence or absence of the above-described annotation. For example, when a certain amount of annotated real data can be obtained, the learning process using non-annotated real data may be omitted.

[0044] [4. Hardware Configuration Example] FIG. 7 is a diagram showing a hardware configuration example of the learning system 1.

[0045] The learning system 1 can be realized by a computer 1000 as shown in FIG. 7. The computer 1000 includes a processing circuitry 1100, a RAM 1200, a ROM 1300, a secondary storage device 1400, a communication interface 1500, an input / output interface 1600, a display unit 1700, a camera unit 1800, a microphone 1900, and a speaker 2000. Each part of the computer 1000 is connected by a bus 1050.

[0046] The processing circuitry 1100 operates based on a program stored in the ROM 1300 or the secondary storage device 1400 and controls each part. For example, the processing circuitry 1100 expands a program stored in the ROM 1300 or the secondary storage device 1400 into the RAM 1200 and executes processes corresponding to various programs.

[0047] The ROM 1300 stores a boot program such as BIOS (Basic Input Output System) executed by the processing circuitry 1100 when the computer 1000 is started up, and programs dependent on the hardware of the computer 1000.

[0048] [[ID=……]]The secondary storage device 1400 is a computer-readable recording medium that non-temporarily records a program executed by the processing circuitry 1100 and data used by such a program. Specifically, the secondary storage device 1400 is a recording medium that records a program for each process of the learning system 1 according to an embodiment of the present disclosure, which is an example of program data 1450. [[ID=……]] [[ID=……]]

[0049] The communication interface 150 is an interface for the computer 1000 to connect to an external network 1550. For example, the processing circuitry 1100 receives data from other devices or transmits data generated by the processing circuitry 1100 to other devices via the communication interface 1500.

[0050] It should be noted that there are some inaccuracies in the original text you provided. For example, in the description of the communication interface in , it should be "communication interface 1500" instead of "communication interface 150". I have made the correction in the translation.The input / output interface 1600 is an interface for connecting the input / output device 1650 and the computer 1000. For example, the processing circuit 1100 receives data from input devices such as a microphone 1900 or a touch panel via the input / output interface 1600. The processing circuit 1100 also transmits data to output devices such as a display unit 1700 or a speaker 2000 via the input / output interface 1600. The input / output interface 1600 may also function as a media interface for reading programs recorded on a predetermined recording medium (media). Examples of media include optical recording media such as DVDs (Digital Versatile Discs) and PDs (Phase Change Rewritable Discs), magneto-optical recording media such as MOs (Magneto-Optical Discs), tape media, magnetic recording media, or semiconductor memory.

[0051] The display unit 1700 is an interface for displaying information processed by the computer 1000. The display unit 1700 is, for example, a liquid crystal display or an organic electroluminescent display (OLED display). Alternatively, the display unit 1700 may be a touch panel display device or an image projection device.

[0052] The camera unit 1800 is an interface for the computer 1000 to capture images. The microphone 1900 is an interface for the computer 1000 to capture sound. The speaker 2000 is an interface for the computer 1000 to output processed sound. The various parts of the computer 1000 are connected by the bus 1050. Each interface does not necessarily have to be located inside the computer 1000, but may be located outside the computer 1000 via a network or the like. Furthermore, each part of the computer 1000 may be controlled by a circuit different from the processing circuit 1100. For example, the display unit 1700 may be controlled not by the processing circuit 1100, but by a circuit dedicated to display processing provided within the display unit 1700.

[0053] For example, when the computer 1000 functions as a learning system 1 according to the embodiment of this disclosure, the processing circuit 1100 of the computer 1000 functions as various detection units and control units (for example, a data discriminator 20, a learning parameter control unit 30, a learning model 40, and a loss calculation unit 50) included in the learning system 1 by executing a program loaded onto the RAM 1200. In addition, the secondary storage device 1400 stores the information processing program and various data according to this disclosure.

[0054] The processing circuit 1100 reads and executes program data 1450 from the secondary storage device 1400, but as an alternative, these programs may be obtained from other devices via an external network 1550. In other words, the secondary storage device 1400 is not limited to being inside the computer 1000, but may be located outside the computer 1000. The processing circuit 1100 is an example of an integrated circuit, and CPUs, MPUs, GPUs, APUs, ASICs, and FPGAs can all be considered integrated circuits.

[0055] Furthermore, the effects described herein are merely illustrative and not limiting, and other effects may also occur.

[0056] [Note] This technology can also be configured as follows: (1) An information processing system comprising: a data discriminator that discriminates the domain of the training data input for fine tuning; and a training parameter control unit that differentiates the model parameters to be fine-tuned according to the determined domain of the training data. (2) The information processing system according to (1) above, wherein the training parameter control unit determines the model parameters indicating the feature extraction unit as the target of fine tuning in the fine tuning using real data as the training data. (3) The information processing system according to (2) above, wherein the training parameter control unit determines both the model parameters indicating the feature extraction unit and the model parameters indicating the low-rank unit as the target of fine tuning in the fine tuning using real data to which annotations have been added as the training data. (4) The information processing system according to (3) above, wherein the training parameter control unit excludes the model parameters indicating the low-rank unit from the target of fine tuning in the fine tuning using real data to which annotations have not been added as the training data. (5) The information processing system according to (4) above, comprising: (5) an information processing system according to (4) above, wherein an exponential moving average processing is performed on a learning model including the feature extraction unit and the low rank unit to obtain a teacher model; the actual data is input into the teacher model to obtain pseudo-annotations; and a Loss calculation unit calculates the difference between the output of the learning model obtained by inputting the actual data into the learning model and the pseudo-annotations as a loss value. (6) The information processing system according to any one of (3) to (5) above, wherein the learning parameter control unit determines the model parameters representing the low rank unit as the target of the fine tuning in the fine tuning using the synthesized data as the learning data, and excludes the model parameters representing the feature extraction unit from the target of the fine tuning.(7) A computer-based information processing method comprising: (8) In the fine tuning using real data as the training data, the method comprises: determining the domain of the training data input for fine tuning, and making the model parameters subject to fine tuning different according to the determined domain of the training data. (9) In the fine tuning using annotated real data as the training data, the method comprises: determining the model parameters indicating the feature extraction unit as the target of fine tuning. (11) The information processing method according to (10) above, comprising: performing exponential moving average processing on a learning model including the feature extraction unit and the low rank unit to obtain a teacher model; inputting the actual data into the teacher model to obtain pseudo-annotations; and calculating the difference between the output of the learning model obtained by inputting the actual data into the learning model and the pseudo-annotations as a loss value. (12) The information processing method according to any one of (9) to (11) above, wherein in the fine tuning using synthetic data as the learning data, the model parameters representing the low rank unit are determined to be the target of the fine tuning, and the model parameters representing the feature extraction unit are excluded from the target of the fine tuning. (13) A program that causes a computer to determine the domain of the learning data input for fine tuning, and to make the model parameters that are the target of the fine tuning different according to the determined domain of the learning data.

[0057] 1. Learning System (Information Processing System) 20. Data Discriminator 30. Learning Parameter Control Unit 40. Learning Model 41. Feature Extraction Unit 42. Low-Rank Unit 50. Loss Calculation Unit

Claims

1. An information processing system comprising: a data discriminator that determines the domain of the training data input for fine tuning; and a training parameter control unit that changes the model parameters to be fine-tuned according to the determined domain of the training data.

2. The information processing system according to claim 1, wherein the learning parameter control unit determines the model parameters representing the feature extraction unit as the target of the fine tuning in the fine tuning using real data as the learning data.

3. The information processing system according to claim 2, wherein the learning parameter control unit determines both the model parameter indicating the feature extraction unit and the model parameter indicating the low-rank unit as targets for the fine tuning in the fine tuning using the annotated real data as the learning data.

4. The information processing system according to claim 3, wherein the learning parameter control unit, in the fine tuning using the actual data without annotations as the learning data, excludes the model parameters indicating the low rank portion from the target of the fine tuning.

5. The information processing system according to claim 4, further comprising: an exponential moving average processing applied to a learning model including the feature extraction unit and the low-rank unit to obtain a teacher model; inputting the actual data into the teacher model to obtain pseudo-annotations; and a Loss calculation unit that calculates the difference between the output of the learning model obtained by inputting the actual data into the learning model and the pseudo-annotations as a loss value.

6. The information processing system according to claim 3, wherein the learning parameter control unit, in the fine tuning using the synthesized data as the learning data, determines the model parameters representing the low-rank portion as the target of the fine tuning, and excludes the model parameters representing the feature extraction portion from the target of the fine tuning.

7. A computer-based information processing method comprising: determining the domain of the training data input for fine-tuning; and differentiating the model parameters subject to fine-tuning according to the determined domain of the training data.

8. The information processing method according to claim 7, wherein in the fine tuning using real data as training data, the model parameters indicating the feature extraction unit are determined to be the target of the fine tuning.

9. The information processing method according to claim 8, wherein in the fine tuning using the annotated real data as training data, both the model parameter indicating the feature extraction unit and the model parameter indicating the low-rank unit are determined to be the target of the fine tuning.

10. The information processing method according to claim 9, wherein in the fine tuning using the actual data that has not been annotated as training data, the model parameters that indicate the low rank portion are excluded from the target of the fine tuning.

11. The information processing method according to claim 10, comprising: performing exponential moving average processing on a learning model including the feature extraction unit and the low-rank unit to obtain a teacher model; inputting the actual data into the teacher model to obtain pseudo-annotations; and calculating the difference between the output of the learning model obtained by inputting the actual data into the learning model and the pseudo-annotations as a loss value.

12. The information processing method according to claim 9, wherein, in the fine tuning using the synthesized data as the training data, the model parameters representing the low-rank portion are determined to be the target of the fine tuning, and the model parameters representing the feature extraction portion are excluded from the target of the fine tuning.

13. A program that enables a computer to determine the domain of the training data input for fine tuning, and to change the model parameters targeted for fine tuning according to the determined domain of the training data.

Images