• Genomics & Informatics
• /
• 제17권2호
• /
• pp.20.1-20.5
• /
• 2019

Entity normalization, or entity linking in the general domain, is an information extraction task that aims to annotate/bind multiple words/expressions in raw text with semantic references, such as concepts of an ontology. An ontology consists minimally of a formally organized vocabulary or hierarchy of terms, which captures knowledge of a domain. Presently, machine-learning methods, often coupled with distributional representations, achieve good performance. However, these require large training datasets, which are not always available, especially for tasks in specialized domains. CONTES (CONcept-TErm System) is a supervised method that addresses entity normalization with ontology concepts using small training datasets. CONTES has some limitations, such as it does not scale well with very large ontologies, it tends to overgeneralize predictions, and it lacks valid representations for the out-of-vocabulary words. Here, we propose to assess different methods to reduce the dimensionality in the representation of the ontology. We also propose to calibrate parameters in order to make the predictions more accurate, and to address the problem of out-of-vocabulary words, with a specific method.

• 로봇학회논문지
• /
• 제18권3호
• /
• pp.346-351
• /
• 2023

With the release of numerous open driving datasets, the demand for domain adaptation in perception tasks has increased, particularly when transferring knowledge from rich datasets to novel domains. However, it is difficult to solve the change 1) in the sensor domain caused by heterogeneous LiDAR sensors and 2) in the environmental domain caused by different environmental factors. We overcome domain differences in the semi-supervised setting with 3-stage model parameter training. First, we pre-train the model with the source dataset with object scaling based on statistics of the object size. Then we fine-tine the partially frozen model weights with copy-and-paste augmentation. The 3D points in the box labels are copied from one scene and pasted to the other scenes. Finally, we use the knowledge distillation method to update the student network with a moving average from the teacher network along with a self-training method with pseudo labels. Test-Time Augmentation with varying z values is employed to predict the final results. Our method achieved 3rd place in ECCV 2022 workshop on the 3D Perception for Autonomous Driving challenge.

• 대한원격탐사학회지
• /
• 제39권4호
• /
• pp.395-407
• /
• 2023

Super-resolution (SR) has great significance in image processing because it enables downstream vision tasks with high spatial resolution. Recently, SR studies have adopted deep learning networks and achieved remarkable SR performance compared to conventional example-based methods. Deep-learning-based SR models generally require low-resolution (LR) images and the corresponding high-resolution (HR) images as training dataset. Due to the difficulties in obtaining real-world LR-HR datasets, most SR models have used only HR images and generated LR images with predefined degradation such as bicubic downsampling. However, SR models trained on simple image degradation do not reflect the properties of the images and often result in deteriorated SR qualities when applied to real-world images. In this study, we propose an image degradation model for HR satellite images based on the modulation transfer function (MTF) of an imaging sensor. Because the proposed method determines the image degradation based on the sensor properties, it is more suitable for training SR models on remote sensing images. Experimental results on HR satellite image datasets demonstrated the effectiveness of applying MTF-based filters to construct a more realistic LR-HR training dataset.

• 한국정보통신학회논문지
• /
• 제24권3호
• /
• pp.355-363
• /
• 2020

본 논문에서는 고대 및 중세 시대의 일본 문자에 대한 데이터세트인 Kuzushiji-MNIST와 Kuzushiji-49를 정확하게 분류하기 위한 딥러닝 학습 방법에 대해서 제안한다. 최신의 합성곱 신경망 네트워크들을 분석하여 가장 적합한 네트워크를 선별하고, 이 네트워크를 이용하여 Kuzushiji-MNIST와 Kuzushiji-49 데이터세트를 분류하기 위한 학습 횟수를 선정한다. 또한 Mixup과 Random Erase 등의 학습 방법을 적용하여 높은 정확도를 갖도록 학습을 진행한다. 학습 결과를 살펴보면 MNIST에 대해서는 99.75%, K-MNIST에 대해서는 99.07%, 그리고 K-49에 대해서는 97.56%의 정확도를 보임으로써 제안한 학습 방법이 높은 성능을 보일 수 있음을 증명하였다. 이와 같은 딥러닝 기반의 기술을 통해 동아시아와 서양의 역사, 문학, 그리고 문화를 연구하는 다양한 연구자들에게 좋은 연구 기반을 제공할 것으로 사료된다.

• Journal of Information Processing Systems
• /
• 제18권5호
• /
• pp.628-636
• /
• 2022

Automatic program repair (APR) techniques focus on automatically repairing bugs in programs and providing correct patches for developers, which have been investigated for decades. However, most studies have limitations in repairing complex bugs. To overcome these limitations, we developed an approach that augments datasets by utilizing token tagging and applying machine learning techniques for APR. First, to alleviate the data insufficiency problem, we augmented datasets by extracting all the methods (buggy and non-buggy methods) in the program source code and conducting token tagging on non-buggy methods. Second, we fed the preprocessed code into the model as an input for training. Finally, we evaluated the performance of the proposed approach by comparing it with the baselines. The results show that the proposed approach is efficient for augmenting datasets using token tagging and is promising for APR.

• ETRI Journal
• /
• 제41권3호
• /
• pp.358-370
• /
• 2019

The main aim of this study is to select the optimal set of genes from microarray cancer datasets that contribute to the prediction of specific cancer types. This study proposes the enhancement of the feature selection filter algorithm based on Joe's normalized mutual information and its use for gene selection. The proposed algorithm is implemented and evaluated on seven benchmark microarray cancer datasets, namely, central nervous system, leukemia (binary), leukemia (3 class), leukemia (4 class), lymphoma, mixed lineage leukemia, and small round blue cell tumor, using five well-known classifiers, including the naive Bayes, radial basis function network, instance-based classifier, decision-based table, and decision tree. An average increase in the prediction accuracy of 5.1% is observed on all seven datasets averaged over all five classifiers. The average reduction in training time is 2.86 seconds. The performance of the proposed method is also compared with those of three other popular mutual information-based feature selection filters, namely, information gain, gain ratio, and symmetric uncertainty. The results are impressive when all five classifiers are used on all the datasets.

• ETRI Journal
• /
• 제43권4호
• /
• pp.630-639
• /
• 2021

Autonomous driving requires a computerized perception of the environment for safety and machine-learning evaluation. Recognizing semantic information is difficult, as the objective is to instantly recognize and distinguish items in the environment. Training a model with real-time semantic capability and high reliability requires extensive and specialized datasets. However, generalized datasets are unavailable and are typically difficult to construct for specific tasks. Hence, a light detection and ranging semantic dataset suitable for semantic simultaneous localization and mapping and specialized for autonomous driving is proposed. This dataset is provided in a form that can be easily used by users familiar with existing two-dimensional image datasets, and it contains various weather and light conditions collected from a complex and diverse practical setting. An incremental and suggestive annotation routine is proposed to improve annotation efficiency. A model is trained to simultaneously predict segmentation labels and suggest class-representative frames. Experimental results demonstrate that the proposed algorithm yields a more efficient dataset than uniformly sampled datasets.

• International Journal of Computer Science & Network Security
• /
• 제22권3호
• /
• pp.29-36
• /
• 2022

In deep learning classification tasks, most models frequently assume that all labels are available for the training datasets. As such strategies to learn new concepts from unlabeled datasets are scarce. In fingerprint classification tasks, most of the fingerprint datasets are labelled using the subject/individual and fingerprint datasets labelled with finger type classes are scarce. In this paper, authors have developed approaches of classifying fingerprint images using the majorly known fingerprint classes. Our study provides a flexible method to learn new classes of fingerprints. Our classifier model combines both the clustering technique and use of deep learning to cluster and hence label the fingerprint images into appropriate classes. The K means clustering strategy explores the label uncertainty and high-density regions from unlabeled data to be clustered. Using similarity index, five clusters are created. Deep learning is then used to train a model using a publicly known fingerprint dataset with known finger class types. A prediction technique is then employed to predict the classes of the clusters from the trained model. Our proposed model is better and has less computational costs in learning new classes and hence significantly saving on labelling costs of fingerprint images.

• 드라이브 ㆍ 컨트롤
• /
• 제19권4호
• /
• pp.104-109
• /
• 2022

For autonomous driving research using AI, datasets collected from road environments play an important role. In other countries, various datasets such as CityScapes, A2D2, and BDD have already been released, but datasets suitable for the domestic road environment still need to be provided. This paper analyzed and verified the dataset reflecting the Korean driving environment. In order to verify the training dataset, the class imbalance was confirmed by comparing the number of pixels and instances of the dataset. A similar A2D2 dataset was trained with the same deep learning model, ConvNeXt, to compare and verify the constructed dataset. IoU was compared for the same class between two datasets with ConvNeXt and mIoU was compared. In this paper, it was confirmed that the collected dataset reflecting the driving environment of Korea is suitable for learning.

• 인터넷정보학회논문지
• /
• 제24권4호
• /
• pp.127-135
• /
• 2023

본 논문에서는 금융 뉴스 데이터로 추가적인 사전 학습이 진행된 BERT 기반 모델인 FinBERT 모델을 사용하여 금융 영역에서 감성 분석 시 학습시킬 데이터와 그에 맞는 하이퍼파라미터를 찾는 방법을 소개한다. 우리의 목표는 다양한 데이터 세트를 활용하고 하이퍼파라미터를 미세 조정하여 정확한 감성 분석을 위해 FinBERT 모델을 가장 잘 활용하는 방법에 대한 포괄적인 가이드를 제공하는 것이다. 이 연구에서는 제안된 FinBERT 모델 미세 조정 접근법의 아키텍처와 워크플로우를 개괄적으로 설명하고, 감성 분석 태스크를 위한 다양한 데이터 세트와 하이퍼파라미터의 성능을 강조한다. 또한, 감성 라벨링 작업에 GPT-3를 사용함으로써 GPT-3가 적절한 라벨러 역할을 하는지에 대한 신뢰성을 검증한다. 결과적으로 미세 조정된 FinBERT 모델이 다양한 데이터 세트에서 우수한 성능을 발휘 한다는 것을 보여주었고, 각 데이터 세트에 대해 전반적으로 우수한 성능을 보이는 학습률 5e-5와 배치 크기 64의 최적의 조합을 찾았다. 또 일반 도메인의 뉴스보다 일반 도메인의 트위터 데이터 세트에서 성능이 크게 향상됨을 기반으로 금융 뉴스 데이터만으로만 추가적으로 학습시키는 FinBERT 모델에 대한 의구심을 제시한다. 이를 통해 FinBERT 모델에 대한 최적의 접근 방식을 결정하는 복잡한 프로세스를 간소화하고 금융 분야 감성 분석 모델을 위한 추가적인 학습 데이터 세트와 미세 조정 시 하이퍼파라미터 선정에 대한 가이드라인을 제시한다.