• The Journal of the Acoustical Society of Korea
• /
• v.41 no.6
• /
• pp.647-654
• /
• 2022

Although the fields in which is utilized children's speech recognition is on the rise, the lack of quality data is an obstacle to improving children's speech recognition performance. This paper proposes a new method for improving children's speech recognition performance by additionally using adult speech data. The proposed method is a transformer based domain adversarial training using dynamically weighted loss to effectively address the data imbalance gap between age that grows as the amount of adult training data increases. Specifically, the degree of class imbalance in the mini-batch during training was quantified, and the loss function was defined and used so that the smaller the data, the greater the weight. Experiments validate the utility of proposed domain adversarial training following asymmetry between adults and children training data. Experiments show that the proposed method has higher children's speech recognition performance than traditional domain adversarial training method under all conditions in which asymmetry between age occurs in the training data.

• Journal of information and communication convergence engineering
• /
• v.21 no.2
• /
• pp.139-144
• /
• 2023

Engineers prefer deep neural networks (DNNs) for solving computer vision problems. However, DNNs pose two major problems. First, neural networks require large amounts of well-labeled data for training. Second, the covariate shift problem is common in computer vision problems. Domain adaptation has been proposed to mitigate this problem. Recent work on adversarial-learning-based unsupervised domain adaptation (UDA) has explained transferability and enabled the model to learn robust features. Despite this advantage, current methods do not guarantee the distinguishability of the latent space unless they consider class-aware information of the target domain. Furthermore, source and target examples alone cannot efficiently extract domain-invariant features from the encoded spaces. To alleviate the problems of existing UDA methods, we propose the mixup regularization in adversarial discriminative domain adaptation (ADDA) method. We validated the effectiveness and generality of the proposed method by performing experiments under three adaptation scenarios: MNIST to USPS, SVHN to MNIST, and MNIST to MNIST-M.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.468-475
• /
• 2020

A good acoustic word embedding that can well express the characteristics of word plays an important role in wake-up word detection (WWD). However, the representation ability of acoustic word embedding may be weakened due to various types of environmental noise occurred in the place where WWD works, causing performance degradation. In this paper, we proposed triplet loss based Domain Adversarial Training (tDAT) mitigating environmental factors that can affect acoustic word embedding. Through experiments in noisy environments, we verified that the proposed method effectively improves the conventional DAT approach, and checked its scalability by combining with other method proposed for robust WWD.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.10
• /
• pp.3211-3229
• /
• 2022

To solve the problems existing in the process of Weibo disaster rumor recognition, such as lack of corpus, poor text standardization, difficult to learn semantic information, and simple semantic features of disaster rumor text, this paper takes Sina Weibo as the data source, constructs a dataset for Weibo disaster rumor recognition, and proposes a deep learning model BERT_AT_Stacked LSTM for Weibo disaster rumor recognition. First, add adversarial disturbance to the embedding vector of each word to generate adversarial samples to enhance the features of rumor text, and carry out adversarial training to solve the problem that the text features of disaster rumors are relatively single. Second, the BERT part obtains the word-level semantic information of each Weibo text and generates a hidden vector containing sentence-level feature information. Finally, the hidden complex semantic information of poorly-regulated Weibo texts is learned using a Stacked Long Short-Term Memory (Stacked LSTM) structure. The experimental results show that, compared with other comparative models, the model in this paper has more advantages in recognizing disaster rumors on Weibo, with an F1_Socre of 97.48%, and has been tested on an open general domain dataset, with an F1_Score of 94.59%, indicating that the model has better generalization.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.3
• /
• pp.771-791
• /
• 2022

Protein-protein interaction (PPI) extraction from original text is important for revealing the molecular mechanism of biological processes. With the rapid growth of biomedical literature, manually extracting PPI has become more time-consuming and laborious. Therefore, the automatic PPI extraction from the raw literature through natural language processing technology has attracted the attention of the majority of researchers. We propose a PPI extraction model based on the large pre-trained language model and adversarial training. It enhances the learning of semantic and syntactic features using BioBERT pre-trained weights, which are built on large-scale domain corpora, and adversarial perturbations are applied to the embedding layer to improve the robustness of the model. Experimental results showed that the proposed model achieved the highest F1 scores (83.93% and 90.31%) on two corpora with large sample sizes, namely, AIMed and BioInfer, respectively, compared with the previous method. It also achieved comparable performance on three corpora with small sample sizes, namely, HPRD50, IEPA, and LLL.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2020.01a
• /
• pp.11-14
• /
• 2020

The importance of data in the development of deep learning is very high. Data with high morphological features are usually utilized in the domains where careful lens calibrations are needed by a human to capture those data. Synthesis of high morphological data for that domain can be a great asset to improve the classification accuracy of systems in the field. Unsupervised learning can be employed for this task. Generating photo-realistic objects of interest has been massively studied after Generative Adversarial Network (GAN) was introduced. In this paper, we propose Morpho-GAN, a method that unifies several GAN techniques to generate quality data of high morphology. Our method introduces a new suitable training objective in the discriminator of GAN to synthesize images that follow the distribution of the original dataset. The results demonstrate that the proposed method can generate plausible data as good as other modern baseline models while taking a less complex during training.

• International Journal of Advanced Culture Technology
• /
• v.6 no.4
• /
• pp.275-283
• /
• 2018

Cancer show distinct pattern of gene expression when it is compared to normal. This difference results malignant characteristic of cancer. Many cancer drugs are targeting this difference so that it can selectively kill cancer cells. One of the recent demand for personalized treating cancer is retrieving normal tissue from a patient so that the gene expression difference between cancer and normal be assessed. However, in most clinical situation it is hard to retrieve normal tissue from a patient. This is because biopsy of normal tissues may cause damage to the organ function or a risk of infection or side effect what a patient to take. Thus, there is a challenge to estimate normal cell's gene expression where cancers are originated from without taking additional biopsy. In this paper, we propose in-silico based prediction of normal cell's gene expression from gene expression data of a tumor sample. We call this challenge as reverting the cancer into normal. We divided this challenge into two parts. The first part is making a generator that is able to fool a pretrained discriminator. Pretrained discriminator is from the training of public data (9,601 cancers, 7,240 normals) which shows 0.997 of accuracy to discriminate if a given gene expression pattern is cancer or normal. Deceiving this pretrained discriminator means our method is capable of generating very normal-like gene expression data. The second part of the challenge is to address whether generated normal is similar to true reverse form of the input cancer data. We used, cycle-consistent adversarial networks to approach our challenges, since this network is capable of translating one domain to the other while maintaining original domain's feature and at the same time adding the new domain's feature. We evaluated that, if we put cancer data into a cycle-consistent adversarial network, it could retain most of the information from the input (cancer) and at the same time change the data into normal. We also evaluated if this generated gene expression of normal tissue would be the biological reverse form of the gene expression of cancer used as an input.

• Journal of KIBIM
• /
• v.9 no.1
• /
• pp.42-51
• /
• 2019

Maintenance of aging structures has attracted societal attention. Maintenance of the aging structure can be efficiently performed with a digital twin. In order to maintain the structure based on the digital twin, it is required to accurately detect the damage of the structure. Meanwhile, deep learning-based damage detection approaches have shown good performance for detecting damage of structures. However, in order to develop such deep learning-based damage detection approaches, it is necessary to use a large number of data before and after damage, but there is a problem that the amount of data before and after the damage is unbalanced in reality. In order to solve this problem, this study proposed a method based on Generative adversarial network, one of Generative Model, for generating acceleration data usually used for damage detection approaches. As results, it is confirmed that the acceleration data generated by the GAN has a very similar pattern to the acceleration generated by the simulation with structural analysis software. These results show that not only the pattern of the macroscopic data but also the frequency domain of the acceleration data can be reproduced. Therefore, these findings show that the GAN model can analyze complex acceleration data on its own, and it is thought that this data can help training of the deep learning-based damage detection approaches.

• Korean Journal of Remote Sensing
• /
• v.38 no.2
• /
• pp.199-213
• /
• 2022

The unsupervised domain adaptation can solve the impractical issue of repeatedly collecting high-quality training data every year for annual crop classification. This study evaluates the applicability of deep learning-based unsupervised domain adaptation models for crop classification. Three unsupervised domain adaptation models including a deep adaptation network (DAN), a deep reconstruction-classification network, and a domain adversarial neural network (DANN) are quantitatively compared via a crop classification experiment using unmanned aerial vehicle images in Hapcheon-gun and Changnyeong-gun, the major garlic and onion cultivation areas in Korea. As source baseline and target baseline models, convolutional neural networks (CNNs) are additionally applied to evaluate the classification performance of the unsupervised domain adaptation models. The three unsupervised domain adaptation models outperformed the source baseline CNN, but the different classification performances were observed depending on the degree of inconsistency between data distributions in source and target images. The classification accuracy of DAN was higher than that of the other two models when the inconsistency between source and target images was low, whereas DANN has the best classification performance when the inconsistency between source and target images was high. Therefore, the extent to which data distributions of the source and target images match should be considered to select the best unsupervised domain adaptation model to generate reliable classification results.

• Journal of the Korea Convergence Society
• /
• v.12 no.10
• /
• pp.63-70
• /
• 2021

Generating a super-resolution meteological data by using a high-resolution deep neural network can provide precise research and useful real-life services. We propose a new technique of generating improved training data for super-resolution deep neural networks. To generate high-resolution meteorological data with domain specific knowledge, Lambert conformal conic projection and objective analysis were applied based on observation data and ERA5 reanalysis field data of specialized institutions. As a result, temperature and humidity analysis data based on domain specific knowledge showed improved RMSE by up to 42% and 46%, respectively. Next, a super-resolution generative adversarial network (SRGAN) which is one of the aritifial intelligence techniques was used to automate the manual data generation technique using damain specific techniques as described above. Experiments were conducted to generate high-resolution data with 1 km resolution from global model data with 10 km resolution. Finally, the results generated with SRGAN have a higher resoltuion than the global model input data, and showed a similar analysis pattern to the manually generated high-resolution analysis data, but also showed a smooth boundary.