• Journal of Korea Multimedia Society
• /
• v.24 no.7
• /
• pp.868-879
• /
• 2021

PCB (Printed Circuit Board) inspection using a deep learning model requires a large amount of data and storage. When the amount of stored data increases, problems such as learning time and insufficient storage space occur. In this study, the existing object detection model is changed to a continual learning model to enable the recognition and classification of PCB components that are constantly increasing. By changing the structure of the object detection model to a knowledge distillation model, we propose a method that allows knowledge distillation of information on existing classified parts while simultaneously learning information on new components. In classification scenario, the transfer learning model result is 75.9%, and the continual learning model proposed in this study shows 90.7%.

• Journal of Korea Multimedia Society
• /
• v.25 no.4
• /
• pp.600-609
• /
• 2022

In supervised learning, labeling of all data is essential, and in particular, in the case of object detection, all objects belonging to the image and to be learned have to be labeled. Due to this problem, continual learning has recently attracted attention, which is a way to accumulate previous learned knowledge and minimize catastrophic forgetting. In this study, a continaul learning model is proposed that accumulates previously learned knowledge and enables learning about new objects. The proposed method is applied to CenterNet, which is a object detection model of anchor-free manner. In our study, the model is applied the knowledge distillation algorithm to be enabled continual learning. In particular, it is assumed that all output layers of the model have to be distilled in order to be most effective. Compared to LWF, the proposed method is increased by 23.3%p mAP in 19+1 scenarios, and also rised by 28.8%p in 15+5 scenarios.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.609-611
• /
• 2021

기계 독해는 주어진 지문 내에서 질문에 대한 답을 기계가 찾아 답하는 문제이다. 딥러닝에서는 여러 데이터셋을 학습시킬 때에 이전에 학습했던 데이터의 weight값이 점차 사라지고 사라진 데이터에 대해 테스트 하였을때 성능이 떨어진 결과를 보인다. 이를 과거에 학습시킨 데이터의 정보를 계속 가진 채로 새로운 데이터를 학습할 수 있는 Continual learning을 통해 해결할 수 있고, 본 논문에서는 이 방법을 MRC에 적용시켜 학습시킨 후 한국어 자연어처리 Task인 Korquad 1.0의 MRC dev set을 통해 성능을 측정하였다. 세 개의 데이터셋중에서 랜덤하게 5만개를 추출하여 10stage를 학습시킨 50K 모델에서 추가로 Continual Learning의 Learning without Forgetting를 사용하여 학습시킨 50K-LWF 모델이 F1 92.57, EM 80.14의 성능을 보였고, BERT 베이스라인 모델의 성능 F1 91.68, EM 79.92에 비교하였을 때 F1, EM 각 0.89, 0.22의 향상이 있었다.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.320-323
• /
• 2022

상호참조해결은 문서에서 명사, 대명사, 명사구 등의 멘션 후보를 식별하고 동일한 개체를 의미하는 멘션들을 찾아 그룹화하는 태스크이다. 딥러닝 기반의 한국어 상호참조해결 연구들에서는 BERT를 이용하여 단어의 문맥 표현을 얻은 후 멘션 탐지와 상호참조해결을 동시에 수행하는 End-to-End 모델이 주로 연구가 되었으며, 최근에는 스팬 표현을 사용하지 않고 시작과 끝 표현식을 통해 상호참조해결을 빠르게 수행하는 Start-to-End 방식의 한국어 상호참조해결 모델이 연구되었다. 최근에 한국어 상호참조해결을 위해 구축된 ETRI 데이터셋은 WIKI, QA, CONVERSATION 등 다양한 도메인으로 이루어져 있으며, 신규 도메인의 데이터가 추가될 경우 신규 데이터가 추가된 전체 학습데이터로 모델을 다시 학습해야 하며, 이때 많은 시간이 걸리는 문제가 있다. 본 논문에서는 이러한 상호참조해결 모델의 도메인 적응에 Continual learning을 적용해 각기 다른 도메인의 데이터로 모델을 학습 시킬 때 이전에 학습했던 정보를 망각하는 Catastrophic forgetting 현상을 억제할 수 있음을 보인다. 또한, Continual learning의 성능 향상을 위해 2가지 Transfer Techniques을 함께 적용한 실험을 진행한다. 실험 결과, 본 논문에서 제안한 모델이 베이스라인 모델보다 개발 셋에서 3.6%p, 테스트 셋에서 2.1%p의 성능 향상을 보였다.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.19 no.3
• /
• pp.151-158
• /
• 2024

Although on-device artificial intelligence (AI) has gained attention to diagnosing machine faults in real time, most previous studies did not consider the model retraining and redeployment processes that must be performed in real-world industrial environments. Our study addresses this challenge by proposing an on-device AI-based real-time machine fault diagnosis system that utilizes continual learning. Our proposed system includes a lightweight convolutional neural network (CNN) model, a continual learning algorithm, and a real-time monitoring service. First, we developed a lightweight 1D CNN model to reduce the cost of model deployment and enable real-time inference on the target edge device with limited computing resources. We then compared the performance of five continual learning algorithms with three public bearing fault datasets and selected the most effective algorithm for our system. Finally, we implemented a real-time monitoring service using an open-source data visualization framework. In the performance comparison results between continual learning algorithms, we found that the replay-based algorithms outperformed the regularization-based algorithms, and the experience replay (ER) algorithm had the best diagnostic accuracy. We further tuned the number and length of data samples used for a memory buffer of the ER algorithm to maximize its performance. We confirmed that the performance of the ER algorithm becomes higher when a longer data length is used. Consequently, the proposed system showed an accuracy of 98.7%, while only 16.5% of the previous data was stored in memory buffer. Our lightweight CNN model was also able to diagnose a fault type of one data sample within 3.76 ms on the Raspberry Pi 4B device.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.1
• /
• pp.27-38
• /
• 2023

In this paper, we propose a conversational AI agent based on continual learning that can continuously learn and grow with new data over time. A continual learning-based conversational AI agent consists of three main components: Task manager, User attribute extraction, and Auto-growing knowledge graph. When a task manager finds new data during a conversation with a user, it creates a new task with previously learned knowledge. The user attribute extraction model extracts the user's characteristics from the new task, and the auto-growing knowledge graph continuously learns the new external knowledge. Unlike the existing conversational AI agents that learned based on a limited dataset, our proposed method enables conversations based on continuous user attribute learning and knowledge learning. A conversational AI agent with continual learning technology can respond personally as conversations with users accumulate. And it can respond to new knowledge continuously. This paper validate the possibility of our proposed method through experiments on performance changes in dialogue generation models over time.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.4
• /
• pp.143-152
• /
• 2021

It is very important to learn behavioral policies that allow multiple agents to work together organically for common goals in various real-world applications. In this multi-agent reinforcement learning (MARL) environment, most existing studies have adopted centralized training with decentralized execution (CTDE) methods as in effect standard frameworks. However, this multi-agent reinforcement learning method is difficult to effectively cope with in a dynamic environment in which new environmental changes that are not experienced during training time may constantly occur in real life situations. In order to effectively cope with this dynamic environment, this paper proposes a novel multi-agent reinforcement learning system, C-COMA. C-COMA is a continual learning model that assumes actual situations from the beginning and continuously learns the cooperative behavior policies of agents without dividing the training time and execution time of the agents separately. In this paper, we demonstrate the effectiveness and excellence of the proposed model C-COMA by implementing a dynamic mini-game based on Starcraft II, a representative real-time strategy game, and conducting various experiments using this environment.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.19 no.2
• /
• pp.91-99
• /
• 2024

Collision avoidance system is important to improve the robustness and functional safety of autonomous vehicles. This paper proposes an object-level distance estimation method to develop a collision avoidance system, and it is applied to golfcarts utilized in country club environments. To improve the detection accuracy, we continually trained an object detection model based on pseudo labels generated by a pre-trained detector. Moreover, we propose object-aware depth estimation (OADE) method which trains a depth model focusing on object regions. In the OADE algorithm, we generated dense depth information for object regions by utilizing detection results and sparse LiDAR points, and it is referred to as object-aware LiDAR projection (OALP). By using the OALP maps, a depth estimation model was trained by backpropagating more gradients of the loss on object regions. Experiments were conducted on our custom dataset, which was collected for the travel distance of 22 km on 54 holes in three country clubs under various weather conditions. The precision and recall rate were respectively improved from 70.5% and 49.1% to 95.3% and 92.1% after the continual learning with pseudo labels. Moreover, the OADE algorithm reduces the absolute relative error from 4.76% to 4.27% for estimating distances to obstacles.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.3
• /
• pp.347-354
• /
• 2022

To reduce forgetfulness in continuous learning, in this paper, we propose an improved LwF model based on the knowledge transfer method, and we show its effectiveness by experiment. In LwF, if the domain of the learned data is different or the complexity of the data is different, the previously learned results are inaccurate due to forgetting. In particular, when learning continues from complex data to simple data, the phenomenon tends to get worse. In this paper, to ensure that the previous learning results are sufficiently transferred to the LwF model, we apply the knowledge transfer method to LwF, and propose an algorithm for efficient use. As a result, the forgetting phenomenon was reduced by an average of 8% compared to the existing LwF results, and it was effective even when the learning task became long. In particular, when complex data was first learned, the efficiency was improved more than 30% compared to LwF.

• Electronics and Telecommunications Trends
• /
• v.39 no.5
• /
• pp.12-20
• /
• 2024

As more devices are used across various industries and their performance improves, artificial intelligence applications are being increasingly adopted. Hence, the rapid development of neural networks suitable for diverse devices can determine the competitiveness of companies. Machine learning operations (MLOps), which constitute a framework that supports neural network generation and its immediate application to devices, have become necessary for the development of artificial intelligence. Currently, most MLOps are provided by major companies such as Google, Amazon, and Microsoft, which provide cloud services supported by large-scale computing power. In addition, various services are provided by the open-source project Kubeflow. We examine basic concepts and technology trends in MLOps and unveil additional functions required in industry.