• International conference on construction engineering and project management
• /
• 2024.07a
• /
• pp.1284-1284
• /
• 2024

Leveraging large language models and safety accident report data has unique potential for analyzing construction accidents, including the classification of accident types, injured parts, and work processes, using unstructured free text accident scenarios. We previously proposed a novel approach that harnesses the power of fine-tuned Generative Pre-trained Transformer to classify 6 types of construction accidents (caught-in-between, cuts, falls, struck-by, trips, and other) with an accuracy of 82.33%. Furthermore, we proposed a novel methodology, saliency visualization, to discern which words are deemed important by black box models within a sentence associated with construction accidents. It helps understand how individual words in an input sentence affect the final output and seeks to make the model's prediction accuracy more understandable and interpretable for users. This involves deliberately altering the position of words within a sentence to reveal their specific roles in shaping the overall output. However, the validation of saliency visualization results remains insufficient and needs further analysis. In this context, this study aims to qualitatively validate the effectiveness of saliency visualization methods. In the exploration of saliency visualization, the elements with the highest importance scores were qualitatively validated against the construction accident risk factors (e.g., "the 4m pipe," "ear," "to extract staircase") emerging from Construction Safety Management's Integrated Information data scenarios provided by the Ministry of Land, Infrastructure, and Transport, Republic of Korea. Additionally, construction accident precursors (e.g., "grinding," "pipe," "slippery floor") identified from existing literature, which are early indicators or warning signs of potential accidents, were compared with the words with the highest importance scores of saliency visualization. We observed that the words from the saliency visualization are included in the pre-identified accident precursors and risk factors. This study highlights how employing saliency visualization enhances the interpretability of models based on large language processing, providing valuable insights into the underlying causes driving accident predictions.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2021.07a
• /
• pp.395-398
• /
• 2021

최근 방대한 양의 텍스트 정보가 인터넷에 유통되면서 정보의 핵심 내용을 파악하기가 더욱 어려워졌으며, 이로 인해 자동으로 텍스트를 요약하려는 연구가 활발하게 이루어지고 있다. 텍스트 자동 요약을 위한 다양한 기법 중 특히 트랜스포머(Transformer) 기반의 모델은 추상 요약(Abstractive Summarization) 과제에서 매우 우수한 성능을 보이며, 해당 분야의 SOTA(State of the Art)를 달성하고 있다. 하지만 트랜스포머 모델은 매우 많은 수의 매개변수들(Parameters)로 구성되어 있어서, 충분한 양의 데이터가 확보되지 않으면 이들 매개변수에 대한 충분한 학습이 이루어지지 않아서 양질의 요약문을 생성하기 어렵다는 한계를 갖는다. 이러한 한계를 극복하기 위해 본 연구는 소량의 데이터가 주어진 환경에서도 양질의 요약문을 생성할 수 있는 문서 요약 방법론을 제안한다. 구체적으로 제안 방법론은 한국어 사전학습 언어 모델인 KoBERT의 임베딩 행렬을 트랜스포머 모델에 적용하는 방식으로 문서 요약을 수행하며, 제안 방법론의 우수성은 Dacon 한국어 문서 생성 요약 데이터셋에 대한 실험을 통해 ROUGE 지표를 기준으로 평가하였다.

• Journal of Sensor Science and Technology
• /
• v.30 no.5
• /
• pp.290-294
• /
• 2021

Experts have designed popular and successful model architectures, which, however, were not the optimal option for different scenarios. Despite the remarkable performances achieved by deep neural networks, manually designed networks for classification tasks are the backbone of object detection. One major challenge is the ImageNet pre-training of the search space representation; moreover, the searched network incurs huge computational cost. Therefore, to overcome the obstacle of the pre-training process, we introduce a network adaptation technique using a pre-trained backbone model tested on ImageNet. The adaptation method can efficiently adapt the manually designed network on ImageNet to the new object-detection task. Neural architecture search (NAS) is adopted to adapt the architecture of the network. The adaptation is conducted on the MobileNetV2 network. The proposed NAS is tested using SSDLite detector. The results demonstrate increased performance compared to existing network architecture in terms of search cost, total number of adder arithmetics (Madds), and mean Average Precision(mAP). The total computational cost of the proposed NAS is much less than that of the State Of The Art (SOTA) NAS method.

• Journal of Information Processing Systems
• /
• v.18 no.2
• /
• pp.282-291
• /
• 2022

This research applies a pre-trained bidirectional encoder representations from transformers (BERT) handwriting recognition model to predict foreign Korean-language learners' writing scores. A corpus of 586 answers to midterm and final exams written by foreign learners at the Intermediate 1 level was acquired and used for pre-training, resulting in consistent performance, even with small datasets. The test data were pre-processed and fine-tuned, and the results were calculated in the form of a score prediction. The difference between the prediction and actual score was then calculated. An accuracy of 95.8% was demonstrated, indicating that the prediction results were strong overall; hence, the tool is suitable for the automatic scoring of Korean written test answers, including grammatical errors, written by foreigners. These results are particularly meaningful in that the data included written language text produced by foreign learners, not native speakers.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.2
• /
• pp.503-523
• /
• 2022

Nowadays, many attendance applications utilise biometric techniques such as the face, fingerprint, and iris recognition. Biometrics has become ubiquitous in many sectors. Due to the advancement of deep learning algorithms, the accuracy rate of biometric techniques has been improved tremendously. This paper proposes a web-based attendance system that adopts facial recognition using open-source deep learning pre-trained models. Face recognition procedural steps using web technology and database were explained. The methodology used the required pre-trained weight files embedded in the procedure of face recognition. The face recognition method includes two important processes: registration of face datasets and face matching. The extracted feature vectors were implemented and stored in an online database to create a more dynamic face recognition process. Finally, user testing was conducted, whereby users were asked to perform a series of biometric verification. The testing consists of facial scans from the front, right (30 - 45 degrees) and left (30 - 45 degrees). Reported face recognition results showed an accuracy of 92% with a precision of 100% and recall of 90%.

• Smart Media Journal
• /
• v.12 no.6
• /
• pp.35-40
• /
• 2023

In this study paper, we examine pre-trained language models used in Multi-Goal Conversational Recommender Systems (MG-CRS), comparing and analyzing their performances of various pre-trained language models. Specifically, we investigates the impact of the sizes of language models on the performance of MG-CRS. The study targets three types of language models - of BERT, GPT2, and BART, and measures and compares their accuracy in two tasks of 'type prediction' and 'topic prediction' on the MG-CRS dataset, DuRecDial 2.0. Experimental results show that all models demonstrated excellent performance in the type prediction task, but there were notable provide significant performance differences in performance depending on among the models or based on their sizes in the topic prediction task. Based on these findings, the study provides directions for improving the performance of MG-CRS.

• Korean Journal of Radiology
• /
• v.22 no.12
• /
• pp.2073-2081
• /
• 2021

Deep learning-based applications have great potential to enhance the quality of medical services. The power of deep learning depends on open databases and innovation. Radiologists can act as important mediators between deep learning and medicine by simultaneously playing pioneering and gatekeeping roles. The application of deep learning technology in medicine is sometimes restricted by ethical or legal issues, including patient privacy and confidentiality, data ownership, and limitations in patient agreement. In this paper, we present an open platform, MI2RLNet, for sharing source code and various pre-trained weights for models to use in downstream tasks, including education, application, and transfer learning, to encourage deep learning research in radiology. In addition, we describe how to use this open platform in the GitHub environment. Our source code and models may contribute to further deep learning research in radiology, which may facilitate applications in medicine and healthcare, especially in medical imaging, in the near future. All code is available at https://github.com/mi2rl/MI2RLNet.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.399-401
• /
• 2022

A definitive diagnosis to classify the breast malignancy status may be achieved by microscopic analysis using surgical open biopsy. However, this procedure requires experts in the specializing of histopathological image analysis directing to time-consuming and high cost. To overcome these issues, deep learning is considered practically efficient to categorize breast cancer into benign and malignant from histopathological images in order to assist pathologists. This study presents a pre-trained convolutional neural network model architecture with a 100% fine-tuning scheme and Adagrad optimizer to classify the breast cancer histopathological images into benign and malignant using a 40× magnification BreaKHis dataset. The pre-trained architecture was constructed using the InceptionResNetV2 model to generate a modified InceptionResNetV2 by substituting the last layer with dense and dropout layers. The results by demonstrating training loss of 0.25%, training accuracy of 99.96%, validation loss of 3.10%, validation accuracy of 99.41%, test loss of 8.46%, and test accuracy of 98.75% indicated that the modified InceptionResNetV2 model is reliable to predict the breast malignancy type from histopathological images. Future works are necessary to focus on k-fold cross-validation, optimizer, model, hyperparameter optimization, and classification on 100×, 200×, and 400× magnification.

• International Journal of Advanced Culture Technology
• /
• v.9 no.2
• /
• pp.86-90
• /
• 2021

Sentiment refers to a person's thoughts, opinions, and feelings toward an object. Sentiment analysis is a process of collecting opinions on a specific target and classifying them according to their emotions, and applies to opinion mining that analyzes product reviews and reviews on the web. Companies and users can grasp the opinions of public opinion and come up with a way to do so. Recently, natural language processing models using the Transformer structure have appeared, and Google's BERT is a representative example. Afterwards, various models came out by remodeling the BERT. Among them, the Facebook AI team unveiled the XLM-R (XLM-RoBERTa), an upgraded XLM model. XLM-R solved the data limitation and the curse of multilinguality by training XLM with 2TB or more refined CC (CommonCrawl), not Wikipedia data. This model showed that the multilingual model has similar performance to the single language model when it is trained by adjusting the size of the model and the data required for training. Therefore, in this paper, we study the improvement of Korean sentiment analysis performed using a pre-trained XLM-R model that solved curse of multilinguality and improved performance.

• Imaging Science in Dentistry
• /
• v.53 no.1
• /
• pp.27-34
• /
• 2023

Purpose: The aim of this study was to clarify the influence of training with a different kind of lesion on the performance of a target model. Materials and Methods: A total of 310 patients(211 men, 99 women; average age, 47.9±16.1 years) were selected and their panoramic images were used in this study. We created a source model using panoramic radiographs including mandibular radiolucent cyst-like lesions (radicular cyst, dentigerous cyst, odontogenic keratocyst, and ameloblastoma). The model was simulatively transferred and trained on images of Stafne's bone cavity. A learning model was created using a customized DetectNet built in the Digits version 5.0 (NVIDIA, Santa Clara, CA). Two machines(Machines A and B) with identical specifications were used to simulate transfer learning. A source model was created from the data consisting of ameloblastoma, odontogenic keratocyst, dentigerous cyst, and radicular cyst in Machine A. Thereafter, it was transferred to Machine B and trained on additional data of Stafne's bone cavity to create target models. To investigate the effect of the number of cases, we created several target models with different numbers of Stafne's bone cavity cases. Results: When the Stafne's bone cavity data were added to the training, both the detection and classification performances for this pathology improved. Even for lesions other than Stafne's bone cavity, the detection sensitivities tended to increase with the increase in the number of Stafne's bone cavities. Conclusion: This study showed that using different lesions for transfer learning improves the performance of the model.