• Journal of the Korean Society for Aeronautical & Space Sciences
• /
• v.49 no.10
• /
• pp.883-892
• /
• 2021

This paper provides a process for developing deep learning-based aerial object detection models that can run in realtime on onboard. To improve object detection performance, we pre-process and augment the training data in the training stage. In addition, we perform transfer learning and apply a weighted cross-entropy method to reduce the variations of detection performance for each class. To improve the inference speed, we have generated inference acceleration engines with quantization. Then, we analyze the real-time performance and detection performance on custom aerial image dataset to verify generalization.

• Broadcasting and Media Magazine
• /
• v.25 no.1
• /
• pp.28-41
• /
• 2020

인공지능의 고속화를 위한 인공지능용 혹은 딥러닝용 하드웨어 및 소프트웨어 시스템에 대한 수요가 폭발적으로 증가하고 있다. 또한 딥러닝 모델에 따라 다양한 추론 시스템이 끊임없이 연구되고 소개되고 있다. 최근에는 전세계에서 100개가 넘는 회사들에서 인공지능용 추론 칩을 개발하고 있고, 임베디드 시스템에서 데이터센터 솔루션에 이르기까지 다양한 분야를 위한 것들이 존재한다. 이러한 하드웨어의 개발을 위해서 12개 이상의 소프트웨어 프레임 워크 및 라이브러리가 활용되고 있다. 하드웨어와 소프트웨어가 다양한 만큼 이들을 중립적으로 평가하기가 매우 어려운 실정이다. 따라서 업계 표준의 인공지능을 위한 벤치마킹 및 평가기준이 필요한데, 이러한 요구로 인해 MLPerf 추론이 만들어졌다. MLPerf는 30개 이상의 기업과 200개 이상의 머신러닝 연구자 및 실무자들에 의해 운영되고, 전혀 다른 구조를 갖는 시스템을 비교할 수 있는 일관성 있는 규칙과 방법을 제시한다. MLPerf에 의해 제시된 규칙에 의해 2019년도에 처음으로 다양한 인공지능용 추론 하드웨어가 벤치마킹을 수행했다. 여기에는 14개의 회사에서 600개 이상의 추론 결과를 측정하였으며, 30개가 넘는 시스템이 이러한 추론에 사용되었다. 본 원고에서는 MLPerf의 학습과 추론을 중심으로 하여 최근에 개발된 다양한 회사들의 인공지능용 하드웨어, 즉 가속기 들의 성능을 살펴보고자 한다.

• The Journal of the Convergence on Culture Technology
• /
• v.10 no.1
• /
• pp.471-476
• /
• 2024

With the spread of various smart devices and computing devices, big data generation is occurring widely. Machine learning is an algorithm that performs reasoning by learning data patterns. Among the various machine learning algorithms, the algorithm that attracts attention is deep learning based on neural networks. Deep learning is achieving rapid performance improvement with the release of various applications. Recently, among deep learning algorithms, attempts to analyze data using graph structures are increasing. In this study, we present a graph generation method for transferring to a deep learning network. This paper proposes a method of generalizing node properties and edge weights in the graph generation process and converting them into a structure for deep learning input by presenting a matricization We present a method of applying a linear transformation matrix that can preserve attribute and weight information in the graph generation process. Finally, we present a deep learning input structure of a general graph and present an approach for performance analysis.

• Annual Conference of KIPS
• /
• 2021.11a
• /
• pp.545-547
• /
• 2021

발전시설 장비는 이상이 생기면 큰 경제적 피해를 발생시키기 때문에, 장비의 계통마다 수십만 개의 센서들이 부착되어 장비의 정상 작동 여부를 모니터링 한다. 장비의 이상 감지를 위해서, 최근 활발히 연구되고 있는 딥러닝 등의 기술을 활용한 AI 모델을 생성하여 장비의 고장을 예측한다. AI 모델을 학습하고 추론하기 위해서는 수많은 센서 중에서 AI 모델을 생성할 센서들을 선택하고, 지속적으로 모니터링 되는 값들을 비교하여 이상 감지 여부를 스트리밍 환경에서 추론할 수 있는 센서 빅데이터 질의 처리 및 스트리밍 추론 시스템이 필요하다. 본 논문에서는 AI 모델을 학습하고 스트리밍 추론할 수 있는 빅데이터 질의 처리 시스템을 설계 및 구현한다.

• Journal of Korean Tunnelling and Underground Space Association
• /
• v.26 no.2
• /
• pp.129-152
• /
• 2024

In the application of deep learning object detection via CCTV in tunnels, a large number of false positive detections occur due to the poor environmental conditions of tunnels, such as low illumination and severe perspective effect. This problem directly impacts the reliability of the tunnel CCTV-based accident detection system reliant on object detection performance. Hence, it is necessary to reduce the number of false positive detections while also enhancing the number of true positive detections. Based on a deep learning object detection model, this paper proposes a false positive data training method that not only reduces false positives but also improves true positive detection performance through retraining of false positive data. This paper's false positive data training method is based on the following steps: initial training of a training dataset - inference of a validation dataset - correction of false positive data and dataset composition - addition to the training dataset and retraining. In this paper, experiments were conducted to verify the performance of this method. First, the optimal hyperparameters of the deep learning object detection model to be applied in this experiment were determined through previous experiments. Then, in this experiment, training image format was determined, and experiments were conducted sequentially to check the long-term performance improvement through retraining of repeated false detection datasets. As a result, in the first experiment, it was found that the inclusion of the background in the inferred image was more advantageous for object detection performance than the removal of the background excluding the object. In the second experiment, it was found that retraining by accumulating false positives from each level of retraining was more advantageous than retraining independently for each level of retraining in terms of continuous improvement of object detection performance. After retraining the false positive data with the method determined in the two experiments, the car object class showed excellent inference performance with an AP value of 0.95 or higher after the first retraining, and by the fifth retraining, the inference performance was improved by about 1.06 times compared to the initial inference. And the person object class continued to improve its inference performance as retraining progressed, and by the 18th retraining, it showed that it could self-improve its inference performance by more than 2.3 times compared to the initial inference.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.32 no.5
• /
• pp.817-826
• /
• 2022

AI technology is being successfully introduced in many fields, and models deployed as a service are deployed with black box environment that does not expose the model's information to protect intellectual property rights and data. In a black box environment, attackers try to steal data or parameters used during training by using model output. This paper proposes a method of inferring the type of model to directly find out the composition of layer of the target model, based on the fact that there is no attack to infer the information about the type of model from the deep learning model. With ResNet, VGGNet, AlexNet, and simple convolutional neural network models trained with MNIST datasets, we show that the types of models can be inferred using the output values in the gray box and black box environments of the each model. In addition, we inferred the type of model with approximately 83% accuracy in the black box environment if we train the big and small relationship feature that proposed in this paper together, the results show that the model type can be infrerred even in situations where only partial information is given to attackers, not raw probability vectors.

• Journal of Broadcast Engineering
• /
• v.27 no.4
• /
• pp.511-518
• /
• 2022

Supervised learning based on deep learning has made a leap forward in various application fields. However, many supervised learning methods work under the common assumption that training and test data are extracted from the same distribution. If it deviates from this constraint, the deep learning network trained in the training domain is highly likely to deteriorate rapidly in the test domain due to the distribution difference between domains. Domain adaptation is a methodology of transfer learning that trains a deep learning network to make successful inferences in a label-poor test domain (i.e., target domain) based on learned knowledge of a labeled-rich training domain (i.e., source domain). In particular, the unsupervised domain adaptation technique deals with the domain adaptation problem by assuming that only image data without labels in the target domain can be accessed. In this paper, we explore the unsupervised domain adaptation techniques.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2020.07a
• /
• pp.359-362
• /
• 2020

본 연구에서는 설명 문장 생성을 통한 해석 가능한 시각적 질의응답 모델을 설계하고 학습 방법을 제시한다. 설명 문장은 시각적 질의응답 모델이 응답을 예측하는 데에 필요한 이미지 및 질문 정보와 적절한 논리적인 정보의 조합 및 정답 추론 과정이 함의되어 있을 것으로 기대한다. 설명 문장 생성 과정이 포함된 시각적 질의응답의 기본적인 모델을 기반으로 여러 가지 학습방법을 통해 설명 문장 생성 과정과 응답 예측 과정간의 상호관계를 분석한다. 이러한 상호작용을 적극적으로 활용할 수 있는 보다 개선 시각적 질의응답 모델을 제안한다. 또한 학습한 결과를 바탕으로 설명 문장의 특성을 활용하여 시각적 질의응답 추론 과정을 개선함으로써 시각적 질의응답 모델의 발전 방향을 논의한다. 본 실험을 통해서 응답 예측에 적절한 설명 문장을 제시하는 해석 가능한 시각적 질의응답 모델을 제공한다.

• Journal of Korea Society of Industrial Information Systems
• /
• v.25 no.2
• /
• pp.19-27
• /
• 2020

Recently, rapid and accurate 3D models creation is required in various applications using virtual reality and augmented reality technology. In this paper, we propose an on-site learning based shape deformation method which transforms the clothed 3D human model into the shape of an input point cloud. The proposed algorithm consists of two main parts: one is pre-learning and the other is on-site learning. Each learning consists of encoder, template transformation and decoder network. The proposed network is learned by unsupervised method, which uses the Chamfer distance between the input point cloud form and the template vertices as the loss function. By performing on-site learning on the input point clouds during the inference process, the high accuracy of the inference results can be obtained and presented through experiments.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.419-422
• /
• 2023

자연어 추론은 전제 문장과 가설 문장의 관계를 함의, 중립, 모순으로 분류하는 자연어 처리 태스크이다. 최근 여러 자연어 처리 태스크에서 딥러닝 모델을 이용하는 방법이 우수한 성능을 보이고 있지만, 이는 미세 조정과정에 드는 비용이 많다는 점과 모델 출력의 근거, 과정을 사람이 이해하기 어려운 한계가 있다. 이러한 이유로 최근에는 소량의 입력, 출력 예시를 포함한 프롬프트를 이용한 방법론과 모델 출력에 대한 근거를 생성, 활용하는 방법에 관한 많은 연구가 진행되고 있다. 본 논문에서는 퓨샷 학습 환경의 한국어 자연어 추론 태스크를 위한 세 가지 프롬프트 방법과 이들을 조합하여 적용하는 방법을 제안한다. 이를 통해 '해석 가능성'과 자연어 추론 성능을 모두 향상시킬 수 있음을 보인다.