• Journal of Broadcast Engineering
• /
• v.26 no.7
• /
• pp.844-854
• /
• 2021

Semi-supervised learning (SSL) and few-shot learning (FSL) have shown impressive performance even then the volume of labeled data is very limited. However, SSL and FSL can encounter a significant performance degradation if the diversity gap between the labeled and unlabeled data is high. To reduce this diversity gap, we propose a novel scheme that relies on an autoencoder for generating pseudo examples. Specifically, the autoencoder is trained on a specific class using the available labeled data and the decoder of the trained autoencoder is then used to generate N samples of that specific class based on N random noise, sampled from a standard normal distribution. The above process is repeated for all the classes. Consequently, the generated data reduces the diversity gap and enhances the model performance. Extensive experiments on MNIST and FashionMNIST datasets for SSL and FSL verify the effectiveness of the proposed approach in terms of classification accuracy and robustness against adversarial attacks.

• The Journal of Korea Robotics Society
• /
• v.18 no.3
• /
• 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.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.05a
• /
• pp.553-556
• /
• 2022

이상지질혈증의 발병에 대한 조기 진단 및 관리하는 것은 중요한 문제이다. 이상지질혈증의 진단은 혈액계측 정보 중에서 네 가지 LDL, HDL, TG, 그리고 TC를 이용하여 진단하며, 이상지질혈증 관리를 위해서는 LDL을 추정하는 것이 중요하다. 본 논문에서는 나이, 성별, 그리고 BMI와 같은 신체계측 정보를 학습하여 LDL-콜레스테롤을 예측하기 위한 준지도학습(Semi-supervised learning) 기반 기계학습 방법을 제안한다. 제안 방법은 얕은 학습(Shallow Learning)기반의 MLP(Multi-Layer Perceptron)을 이용하고, 이상지질혈증 진단인자간의 상관관계를 고려하여 신체계측 정보로 예측된 HDL, TG, 그리고 TC을 이용하여 일반적인 기계학습을 이용한 예측방법의 정확도를 개선한다. 즉, 제안방법은 신체계측 정보를 이용하여 혈액계측 정보의 LDL, HDL, TG, 그리고 TC을 각각 예측하고, 신체계측에 혈액계측의 예측 정보를 추가하여 학습한 준지도학습 기반 얕은 네트워크를 설계한다. 실험결과, HDL, TG, 그리고 TC의 혈액예측 정보를 이용한 준지도학습 기반 LDL 예측 정확도는 71.4%로 신체계측 정보만을 이용한 예측 방법의 67.0% 보다 약 4.4% 개선할 수 있음을 확인한다.

• Journal of Broadcast Engineering
• /
• v.27 no.1
• /
• pp.13-19
• /
• 2022

Video Object Segmentation (VOS) task requires an amount of labeled sequence data, which limits the performance of the current VOS methods trained with public datasets. In this paper, we propose two effective data augmentation schemes for VOS. The first augmentation method is to swap the background segment to the background from another image, and the other method is to play the sequence in reverse. The two augmentation schemes for VOS enable the current VOS methods to robustly predict the segmentation labels and improve the performance of VOS.

• Asia pacific journal of information systems
• /
• v.29 no.4
• /
• pp.789-816
• /
• 2019

Text classification is a challenging task, especially when dealing with a huge amount of text data. The performance of a classification model can be varied depending on what type of words contained in the document corpus and what type of features generated for classification. Aside from proposing a new modified version of the existing algorithm or creating a new algorithm, we attempt to modify the use of data. The classifier performance is usually affected by the quality of learning data as the classifier is built based on these training data. We assume that the data from different domains might have different characteristics of noise, which can be utilized in the process of learning the classifier. Therefore, we attempt to enhance the robustness of the classifier by injecting the heterogeneous data artificially into the learning process in order to improve the classification accuracy. Semi-supervised approach was applied for utilizing the heterogeneous data in the process of learning the document classifier. However, the performance of document classifier might be degraded by the unlabeled data. Therefore, we further proposed an algorithm to extract only the documents that contribute to the accuracy improvement of the classifier.

• Proceedings of the Korean Information Science Society Conference
• /
• 2008.06c
• /
• pp.268-273
• /
• 2008

의견(Opinion) 분석은 도전적인 분야로 언어 자원 구축, 문서의 Sentiment 분류, 문장 내의 의견 어구 추출 등의 다양한 문제를 다룬다. 이 중 의견 어구 추출문제는 단순히 문장이나 문서 단위로 분류하는 수준을 뛰어 넘는 문장 내 의견 어구를 추출하는 문제로 최근 많은 관심을 받고 있는 연구 주제이다. 그러나 의견 어구 추출에 대한 기존 연구는 문장 내 의견 어구부분이 태깅(tagging)된 학습 데이터와 의견 어휘 자원을 이용한 지도(Supervised)학습을 이용한 접근이 대부분으로 실제 적용 상의 한계를 갖는다. 본 논문은 문장 내 의견 어구 부분이 태깅된 학습 데이터와 의견 어휘 자원이 없는 환경에서도 문장단위의 극성 정보를 이용하여 의견 어구를 추출하는 부분 지도(Semi-Supervised)학습 장법을 제안한다. 본 논문의 방법은 Baseline에 비하여 정확률(Precision)은 33%, F-Measure는 14% 가량 높은 성능을 냈다.

• Korean Journal of Remote Sensing
• /
• v.34 no.6_2
• /
• pp.1239-1249
• /
• 2018

Optical remote sensing sensors provide visually more familiar images than radar images. However, it is difficult to discriminate sea ice types in optical images using spectral information based machine learning algorithms. This study addresses two topics. First, we propose a semantic segmentation which is a part of the state-of-the-art deep learning algorithms to identify ice types by learning hierarchical and spatial features of sea ice. Second, we propose a new approach by combining of semi-supervised and active learning to obtain accurate and meaningful labels from unlabeled or unseen images to improve the performance of supervised classification for multiple images. Therefore, we successfully added new labels from unlabeled data to automatically update the semantic segmentation model. This should be noted that an operational system to generate ice type products from optical remote sensing data may be possible in the near future.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.2
• /
• pp.751-770
• /
• 2019

Action recognition is an essential task in computer vision due to the variety of prospective applications, such as security surveillance, machine learning, and human-computer interaction. The availability of more video data than ever before and the lofty performance of deep convolutional neural networks also make it essential for action recognition in video. Unfortunately, limited crafted video features and the scarcity of benchmark datasets make it challenging to address the multi-person action recognition task in video data. In this work, we propose a deep convolutional neural network-based Effective Hybrid Learning (EHL) framework for two-person interaction classification in video data. Our approach exploits a pre-trained network model (the VGG16 from the University of Oxford Visual Geometry Group) and extends the Faster R-CNN (region-based convolutional neural network a state-of-the-art detector for image classification). We broaden a semi-supervised learning method combined with an active learning method to improve overall performance. Numerous types of two-person interactions exist in the real world, which makes this a challenging task. In our experiment, we consider a limited number of actions, such as hugging, fighting, linking arms, talking, and kidnapping in two environment such simple and complex. We show that our trained model with an active semi-supervised learning architecture gradually improves the performance. In a simple environment using an Intelligent Technology Laboratory (ITLab) dataset from Inha University, performance increased to 95.6% accuracy, and in a complex environment, performance reached 81% accuracy. Our method reduces data-labeling time, compared to supervised learning methods, for the ITLab dataset. We also conduct extensive experiment on Human Action Recognition benchmarks such as UT-Interaction dataset, HMDB51 dataset and obtain better performance than state-of-the-art approaches.

• Journal of Computing Science and Engineering
• /
• v.7 no.2
• /
• pp.112-121
• /
• 2013

The last years have seen a rise of interest in using electroencephalography-based brain computer interfacing methodology for investigating non-medical questions, beyond the purpose of communication and control. One of these novel applications is to examine how signal quality is being processed neurally, which is of particular interest for industry, besides providing neuroscientific insights. As for most behavioral experiments in the neurosciences, the assessment of a given stimulus by a subject is required. Based on an EEG study on speech quality of phonemes, we will first discuss the information contained in the neural correlate of this judgement. Typically, this is done by analyzing the data along behavioral responses/labels. However, participants in such complex experiments often guess at the threshold of perception. This leads to labels that are only partly correct, and oftentimes random, which is a problematic scenario for using supervised learning. Therefore, we propose a novel supervised-unsupervised learning scheme, which aims to differentiate true labels from random ones in a data-driven way. We show that this approach provides a more crisp view of the brain states that experimenters are looking for, besides discovering additional brain states to which the classical analysis is blind.

• Phonetics and Speech Sciences
• /
• v.12 no.2
• /
• pp.29-37
• /
• 2020

Recently, the neural network-based deep learning algorithm has dramatically improved performance compared to the classical Gaussian mixture model based hidden Markov model (GMM-HMM) automatic speech recognition (ASR) system. In addition, researches on end-to-end (E2E) speech recognition systems integrating language modeling and decoding processes have been actively conducted to better utilize the advantages of deep learning techniques. In general, E2E ASR systems consist of multiple layers of encoder-decoder structure with attention. Therefore, E2E ASR systems require data with a large amount of speech-text paired data in order to achieve good performance. Obtaining speech-text paired data requires a lot of human labor and time, and is a high barrier to building E2E ASR system. Therefore, there are previous studies that improve the performance of E2E ASR system using relatively small amount of speech-text paired data, but most studies have been conducted by using only speech-only data or text-only data. In this study, we proposed a semi-supervised training method that enables E2E ASR system to perform well in corpus in different domains by using both speech or text only data. The proposed method works effectively by adapting to different domains, showing good performance in the target domain and not degrading much in the source domain.