• The Journal of the Acoustical Society of Korea
• /
• v.40 no.6
• /
• pp.578-586
• /
• 2021

We propose a semi-supervised learning method based on Variational AutoEncoder (VAE) and Unsupervised Data Augmentation (UDA) to improve the performance of an end-to-end speech recognizer. In the proposed method, first, the VAE-based augmentation model and the baseline end-to-end speech recognizer are trained using the original speech data. Then, the baseline end-to-end speech recognizer is trained again using data augmented from the learned augmentation model. Finally, the learned augmentation model and end-to-end speech recognizer are re-learned using the UDA-based semi-supervised learning method. As a result of the computer simulation, the augmentation model is shown to improve the Word Error Rate (WER) of the baseline end-to-end speech recognizer, and further improve its performance by combining it with the UDA-based learning method.

• The Korean Journal of Applied Statistics
• /
• v.34 no.2
• /
• pp.167-175
• /
• 2021

Semi-supervised learning uses both labeled data and unlabeled data. Recently consistency regularization is very popular in semi-supervised learning. Unsupervised data augmentation (UDA) that uses unlabeled data augmentation is also based on the consistency regularization. The Kullback-Leibler divergence is used for the loss of unlabeled data and cross-entropy for the loss of labeled data through UDA learning. UDA uses techniques such as training signal annealing (TSA) and confidence-based masking to promote performance. In this study, we propose to use Jensen-Shannon divergence instead of Kullback-Leibler divergence, reverse-TSA and not to use confidence-based masking for performance improvement. Through experiment, we show that the proposed technique yields better performance than those of UDA.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.18 no.2
• /
• pp.494-510
• /
• 2024

Internet users are exposed to sophisticated cyberattacks that intrusion detection systems have difficulty detecting. Therefore, research is increasing on intrusion detection methods that use artificial intelligence technology for detecting novel cyberattacks. Unsupervised learning-based methods are being researched that learn only from normal data and detect abnormal behaviors by finding patterns. This study developed an anomaly-detection method based on unsupervised machines and deep learning for a network intrusion detection system (NIDS). We present a hybrid anomaly detection approach based on unsupervised learning techniques using the autoencoder (AE), Isolation Forest (IF), and Local Outlier Factor (LOF) algorithms. An oversampling approach that increased the detection rate was also examined. A hybrid approach that combined deep learning algorithms and traditional machine learning algorithms was highly effective in setting the thresholds for anomalies without subjective human judgment. It achieved precision and recall rates respectively of 88.2% and 92.8% when combining two AEs, IF, and LOF while using an oversampling approach to learn more unknown normal data improved the detection accuracy. This approach achieved precision and recall rates respectively of 88.2% and 94.6%, further improving the detection accuracy compared with the hybrid method. Therefore, in NIDS the proposed approach provides high reliability for detecting cyberattacks.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.618-620
• /
• 2023

최근 자연어처리 분야에서 의미론적 유사성을 반영하기 위한 대조 학습 (contrastive learning) 관련 연구가 활발히 이뤄지고 있다. 이러한 대조 학습의 핵심은 의미론적으로 가까워져야 하는 쌍과 멀어져야 하는 쌍을 잘 구축하는 것이지만, 기존의 손실 함수는 문장의 상대적인 유사성을 풍부하게 반영하는데 한계가 있다. 이를 해결하기 위해, 이전 연구에서는 삼중 항 손실 함수 (triplet loss)를 도입하였으며, 본 논문에서는 이러한 삼중 항을 구성하기 위해 대조 학습에서의 효과적인 토큰 컷오프(cutoff) 데이터 증강 기법을 제안한다. BERT, RoBERTa 등 널리 활용되는 언어 모델을 이용한 실험을 통해 제안하는 방법의 우수한 성능을 보인다.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2024.01a
• /
• pp.73-75
• /
• 2024

본 논문에서는 밀 수확량을 증가시키기 위한 일반화된 검출 모델을 제안한다. 일반화 성능을 높이기 위해 CutMix 알고리즘으로 데이터를 증식시켰고, 라벨링 되지 않은 데이터를 최대한 활용하기 위해 Fast R-CNN 기반 Pseudo labeling을 사용하였다. 학습의 정확성과 효율성을 높이기 위해 사전에 훈련된 EfficientDet 모델로 학습하였으며, OOF를 이용하여 검증하였다. 최신 객체 검출 모델과 IoU(Intersection over Union)를 이용한 성능 평가 결과, 제안된 모델이 가장 높은 성능을 보이는 것을 확인하였다.

• Progress in Medical Physics
• /
• v.30 no.2
• /
• pp.49-58
• /
• 2019

Purpose: Automated analytical systems have begun to emerge as a database system that enables the scanning of medical images to be performed on computers and the construction of big data. Deep-learning artificial intelligence (AI) architectures have been developed and applied to medical images, making high-precision diagnosis possible. Materials and Methods: For diagnosis, the medical images need to be labeled and standardized. After pre-processing the data and entering them into the deep-learning architecture, the final diagnosis results can be obtained quickly and accurately. To solve the problem of overfitting because of an insufficient amount of labeled data, data augmentation is performed through rotation, using left and right flips to artificially increase the amount of data. Because various deep-learning architectures have been developed and publicized over the past few years, the results of the diagnosis can be obtained by entering a medical image. Results: Classification and regression are performed by a supervised machine-learning method and clustering and generation are performed by an unsupervised machine-learning method. When the convolutional neural network (CNN) method is applied to the deep-learning layer, feature extraction can be used to classify diseases very efficiently and thus to diagnose various diseases. Conclusions: AI, using a deep-learning architecture, has expertise in medical image analysis of the nerves, retina, lungs, digital pathology, breast, heart, abdomen, and musculo-skeletal system.