KR-20260067326-A - SYSTEM, DEVICE, METHOD AND COMPUTER PROGRAM FOR LEARNING MODEL

Abstract

A system according to one embodiment of the present disclosure comprises, in a model learning system, a learning module that learns a first model using a learning dataset; an early learning detection module that detects a first time point at which the first model learns first noise data included in the learning dataset and a second time point at which the first model learns second noise data included in the learning dataset; a learning guidance module that guides learning of the first model after the first time point using a second model generated based on the first model at the first time point; and a model update module that updates the weights of the second model based on the weights of the first model.

Inventors

• 최동민
• 이상빈

Assignees

• 세이지 주식회사

Dates

Publication Date

20260512

Application Date

20251028

Priority Date

20241105

Claims (10)

1. In a model learning system, A training module that trains a first model using a training dataset; An early learning detection module that detects a first time point at which the first model learns first noise data included in the training dataset and a second time point at which the first model learns second noise data included in the training dataset; A learning guidance module that guides learning of the first model after the first time point using a second model generated based on the first model at the first time point; and A system comprising a model update module that updates the weights of the second model based on the weights of the first model.
2. In Article 1, The above early learning detection module is, The first performance indicator is modeled as a first exponential parameter function, and the second performance indicator is modeled as a second exponential parameter function, and A system that detects the first time point based on the first exponential parameter function and detects the second time point based on the second exponential parameter function.
3. In Article 2, The above early learning detection module is, Based on the first exponential parameter function above, the point in time when the deviation of the rate of change of the first performance indicator based on the initial learning epoch exceeds a preset first threshold ratio is detected as the first point in time, and A system that detects the point in time when the deviation of the rate of change of the second performance indicator based on the initial learning epoch exceeds a preset second threshold ratio as the second point in time, based on the second exponential parameter function.
4. In Paragraph 3, The above initial learning epoch is a system that is a learning epoch prior to the above first time point.
5. In Article 1, A system in which the learning module generates the second model by copying the weights of the first model at the first time point.
6. In Article 1, The above learning guidance module is a system that guides learning for the first model using knowledge distillation.
7. In Article 6, A system in which the learning guidance module guides the learning of the first model by adding a Knowledge Distillation Loss that minimizes the difference between the prediction result of the first model and the prediction result of the second model to the Total Loss Function of the first model.
8. In Article 1, The above model update module is a system that updates the second model by reflecting the weights of the first model in the second model using an exponential moving average (EMA) method.
9. In Article 1, A system in which the second model is not trained after the first point in time.
10. In Article 1, The above-mentioned first model is a system that is an object detection model.

Description

Model Learning System, Device, Method and Computer Program The present disclosure relates to a model learning system, apparatus, method, and computer program, and more specifically, to a model learning system, apparatus, method, and computer program for learning an object detection model. Recently, interest in technologies such as artificial intelligence, big data, the Internet of Things, and cloud computing has been on the rise. Here, Artificial Intelligence (AI) refers to a technology capable of reasoning, learning, and acting using data that requires human intelligence or is larger in scale than humans can analyze. With recent improvements in computing efficiency, AI-based computer vision technology is rapidly advancing. Computer vision technology is a technology that extracts meaningful information from images, videos, and other visual inputs using computers and systems, and then performs tasks or makes recommendations based on this information. For example, among computer vision technologies, an object detection model is an artificial intelligence model designed to detect the location of objects within an input image and classify their class (or label). Meanwhile, a training dataset is used to train such an object detection model, and the performance of the model can be significantly influenced by the various noises included in the training dataset. FIG. 1 is a block diagram showing an electronic device according to various embodiments of the present disclosure. FIG. 2 is a block diagram illustrating an object detection framework according to various embodiments of the present disclosure. FIG. 3 is a block diagram illustrating a model learning system according to various embodiments of the present disclosure. FIGS. 4a to 4c are drawings illustrating an example of noise in the training data included in the training dataset. Figure 5 is a diagram illustrating an example of differential early learning. Figure 6 is a diagram illustrating an example of the operation of the model learning system of Figure 3. In describing each drawing, similar reference numerals have been used for similar components. In the attached drawings, the dimensions of the structures are depicted enlarged from their actual size for clarity of the present disclosure. The area and thickness of each component shown in the drawings are depicted for convenience of explanation and are not necessarily limited to the area and thickness of the components depicted in the present disclosure. Terms such as "first," "second," etc., may be used to describe various components, but said components should not be limited by said terms. Such terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present disclosure, the first component may be named the second component, and similarly, the second component may be named the first component. A singular expression includes a plural expression unless the context clearly indicates otherwise. In this disclosure, terms such as “comprising,” “having,” and “consisting of” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should not be understood as precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof. The features of each of the various embodiments of the present disclosure may be combined or combined with one another in part or wholly, and various technical interlocking and operation may be possible, and each embodiment may be implemented independently of one another or together in an interlocking relationship. Hereinafter, various embodiments of the present disclosure will be described in more detail with reference to the attached drawings. Identical reference numerals are used for identical components in the drawings, and redundant descriptions of identical components are omitted. <Apparatus according to embodiments of the present disclosure> FIG. 1 is a block diagram showing an electronic device (100) according to various embodiments of the present disclosure. Referring to FIG. 1, an electronic device (100) according to embodiments of the present disclosure may include a processor (110), a communication module (120), an input module (130), a display module (140), and a memory (150). According to an embodiment, a processor (110), a communication module (120), an input module (130), a display module (140), and a memory (150) included in an electronic device (100) may each be electrically and/or physically connected to each other. Meanwhile, although FIG. 1 is illustrated as having an electronic device (100) that includes a processor (110), a communication module (120), an input module (130), a display module (140), and a memory (150), this is merely illustrative and the embodiments of the present disclosure are not limited thereto. For exampl