• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.149-154
• /
• 2023

본고는 한국어 생성 모델의 독성 텍스트 저항 능력을 검증하기 위해 'CVALUE' 데이터셋에서 추출한 고난도 독성 질문-대답 100쌍을 바탕으로 한국어 생성 모델을 위한 '100 K-Poison' 데이터셋을 시범적으로 구축했다. 이 데이터셋을 토대로 4가지 대표적인 한국어 생성 모델 'ZeroShot TextClassifcation'과 'Text Generation7 실험을 진행함으로써 현재 한국어 생성 모델의 독성 텍스트 식별 및 응답 능력을 종합적으로 고찰했고, 모델 간의 독성 텍스트 저항력 격차 현상을 분석했으며, 앞으로 한국어 생성 모델의 독성 텍스트 식별 및 웅대 성능을 한층 더 강화하기 위한 '이독공독(以毒攻毒)' 학습 전략을 새로 제안하였다.

• Journal of Korean Society of Occupational and Environmental Hygiene
• /
• v.33 no.4
• /
• pp.412-418
• /
• 2023

As advances in artificial intelligence (AI) increasingly approach areas once relegated to the realm of science fiction, there is growing public interest in using these technologies for practical everyday tasks in both the home and the workplace. This paper explores the applications of and implications for of using ChatGPT, a conversational AI model based on GPT-3.5 and GPT-4.0, in the field of occupational health and safety. After gaining over one million users within five days of its launch, ChatGPT has shown promise in addressing issues ranging from emergency response to chemical exposure to recommending personal protective equipment. However, despite its potential usefulness, the integration of AI into scientific work and professional settings raises several concerns. These concerns include the ethical dimensions of recognizing AI as a co-author in academic publications, the limitations and biases inherent in the data used to train these models, legal responsibilities in professional contexts, and potential shifts in employment following technological advances. This paper aims to provide a comprehensive overview of these issues and to contribute to the ongoing dialogue on the responsible use of AI in occupational health and safety.

• Annual Conference of KIPS
• /
• 2023.11a
• /
• pp.908-909
• /
• 2023

Recently, several generative models have emerged and are being used in various industries. Among them, Cycle GAN is still used in various fields such as style transfer, medical care and autonomous driving. In this paper, we propose two methods to improve the performance of these Cycle GAN model. The ReLU activation function previously used in the generator was changed to Leaky ReLU. And a new loss function is proposed that considers the semantic level rather than focusing only on the pixel level through the VGG feature extractor. The proposed model showed quality improvement on the test set in the art domain, and it can be expected to be applied to other domains in the future to improve performance.

• Annual Conference of KIPS
• /
• 2023.11a
• /
• pp.461-464
• /
• 2023

최근 GPT-3 와 LLaMa 같은 생성형 거대 언어모델을 활용한 서비스가 공개되었고, 실제로 많은 사람들이 사용하고 있다. 해당 모델들은 사용자들의 다양한 질문에 대해 유창한 답변을 한다는 이유로 주목받고 있다. 하지만 LLMs 의 답변에는 종종 Inconsistent content 와 non-factual statement 가 존재하며, 이는 사용자들로 하여금 잘못된 정보의 전파 등의 문제를 야기할 수 있다. 이에 논문에서는 동일한 질문에 대한 LLM 의 답변 샘플과 외부 지식을 활용한 Hallucination Detection 방법을 제안한다. 제안한 방법은 동일한 질문에 대한 LLM 의 답변들을 이용해 일관성 점수(Consistency score)를 계산한다. 거기에 외부 지식을 이용한 사실검증을 통해 사실성 점수(Factuality score)를 계산한다. 계산된 일관성 점수와 사실성 점수를 활용하여 문장 수준의 Hallucination Detection 을 가능하게 했다. 실험에는 GPT-3 를 이용하여 WikiBio dataset 에 있는 인물에 대한 passage 를 생성한 데이터셋을 사용하였으며, 우리는 해당 방법을 통해 문장 수준에서의 Hallucination Detection 성능이 baseline 보다 AUC-PR scores 에서 향상됨을 보였다.

• Journal of information and communication convergence engineering
• /
• v.22 no.3
• /
• pp.221-230
• /
• 2024

The aim of this study is to investigate the automatic recognition and analysis of Jeju marine-life images using artificial intelligence (AI) technology. The dataset of marine-life images was prepared using tools such as Python, TensorFlow, and Google Colab (Google Colaboratory). We also developed models by training deep learning AI in image recognition to automatically recognize the species found in these images and extract their associated information, such as taxonomy, characteristics, and distribution. This study is innovative in that it uses deep learning technology combined with imagerecognition technology for marine biodiversity research. In addition, these results will lead to the development of the marine-life industry in Jeju by supporting marine environment monitoring and marine resource conservation. Furthermore, this study is anticipated to contribute to academic advancement, specifically in the study of marine species diversity.

• The Bulletin of The Korean Astronomical Society
• /
• v.46 no.1
• /
• pp.41.2-41.2
• /
• 2021

In this study, we generate He I 1083 nm images from Solar Dynamic Observatory (SDO)/Atmospheric Imaging Assembly (AIA) images using a novel deep learning method (pix2pixHD) based on conditional Generative Adversarial Networks (cGAN). He I 1083 nm images from National Solar Observatory (NSO)/Synoptic Optical Long-term Investigations of the Sun (SOLIS) are used as target data. We make three models: single input SDO/AIA 19.3 nm image for Model I, single input 30.4 nm image for Model II, and double input (19.3 and 30.4 nm) images for Model III. We use data from 2010 October to 2015 July except for June and December for training and the remaining one for test. Major results of our study are as follows. First, the models successfully generate He I 1083 nm images with high correlations. Second, the model with two input images shows better results than those with one input image in terms of metrics such as correlation coefficient (CC) and root mean squared error (RMSE). CC and RMSE between real and AI-generated ones for the model III with 4 by 4 binnings are 0.84 and 11.80, respectively. Third, AI-generated images show well observational features such as active regions, filaments, and coronal holes. This work is meaningful in that our model can produce He I 1083 nm images with higher cadence without data gaps, which would be useful for studying the time evolution of chromosphere and coronal holes.

• Korean Journal of Radiology
• /
• v.24 no.11
• /
• pp.1061-1080
• /
• 2023

Artificial intelligence (AI) in radiology is a rapidly developing field with several prospective clinical studies demonstrating its benefits in clinical practice. In 2022, the Korean Society of Radiology held a forum to discuss the challenges and drawbacks in AI development and implementation. Various barriers hinder the successful application and widespread adoption of AI in radiology, such as limited annotated data, data privacy and security, data heterogeneity, imbalanced data, model interpretability, overfitting, and integration with clinical workflows. In this review, some of the various possible solutions to these challenges are presented and discussed; these include training with longitudinal and multimodal datasets, dense training with multitask learning and multimodal learning, self-supervised contrastive learning, various image modifications and syntheses using generative models, explainable AI, causal learning, federated learning with large data models, and digital twins.

• Smart Structures and Systems
• /
• v.34 no.2
• /
• pp.129-143
• /
• 2024

In recent years, with the vigorous development of visual algorithms, a large amount of research has been conducted on blade surface defect detection methods represented by deep learning. Detection methods based on deep learning models must rely on a large and rich dataset. However, the geographical location and working environment of wind turbines makes it difficult to effectively capture images of blade surface defects, which inevitably hinders visual detection. In response to the challenge of collecting a dataset for surface defects that are difficult to obtain, a multi-class blade surface defect generation method based on the StyleGAN3 (Style Generative Adversarial Networks 3) deep learning model and PBGMs (Physics-Based Graphics Models) method has been proposed. Firstly, a small number of real blade surface defect datasets are trained using the adversarial neural network of the StyleGAN3 deep learning model to generate a large number of high-resolution blade surface defect images. Secondly, the generated images are processed through Matting and Resize operations to create defect foreground images. The blade background images produced using PBGM technology are randomly fused, resulting in a diverse and high-resolution blade surface defect dataset with multiple types of backgrounds. Finally, experimental validation has proven that the adoption of this method can generate images with defect characteristics and high resolution, achieving a proportion of over 98.5%. Additionally, utilizing the EISeg annotation method significantly reduces the annotation time to just 1/7 of the time required for traditional methods. These generated images and annotated data of blade surface defects provide robust support for the detection of blade surface defects.

• Journal of Internet Computing and Services
• /
• v.21 no.4
• /
• pp.9-16
• /
• 2020

In the Image processing and computer vision, the problem of reconstructing from one image to another or generating a new image has been steadily drawing attention as hardware advances. However, the problem of computer-generated images also continues to emerge when viewed with human eyes because it is not natural. Due to the recent active research in deep learning, image generating and improvement problem using it are also actively being studied, and among them, the network called Generative Adversarial Network(GAN) is doing well in the image generating. Various models of GAN have been presented since the proposed GAN, allowing for the generation of more natural images compared to the results of research in the image generating. Among them, pix2pix is a conditional GAN model, which is a general-purpose network that shows good performance in various datasets. pix2pix is based on U-Net, but there are many networks that show better performance among U-Net based networks. Therefore, in this study, images are generated by applying various networks to U-Net of pix2pix, and the results are compared and evaluated. The images generated through each network confirm that the pix2pix model with Attention, R2, and Attention-R2 networks shows better performance than the existing pix2pix model using U-Net, and check the limitations of the most powerful network. It is suggested as a future study.

• Journal of Convergence for Information Technology
• /
• v.12 no.4
• /
• pp.260-269
• /
• 2022

In this paper, We propose an anomaly detection model using deep neural network to automate the identification of outliers of the national air pollution measurement network data that is previously performed by experts. We generated training data by analyzing missing values and outliers of weather data provided by the Institute of Environmental Research and based on the BeatGAN model of the unsupervised learning method, we propose a new model by changing the kernel structure, adding the convolutional filter layer and the transposed convolutional filter layer to improve anomaly detection performance. In addition, by utilizing the generative features of the proposed model to implement and apply a retraining algorithm that generates new data and uses it for training, it was confirmed that the proposed model had the highest performance compared to the original BeatGAN models and other unsupervised learning model like Iforest and One Class SVM. Through this study, it was possible to suggest a method to improve the anomaly detection performance of proposed model while avoiding overfitting without additional cost in situations where training data are insufficient due to various factors such as sensor abnormalities and inspections in actual industrial sites.