• 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.

• International journal of advanced smart convergence
• /
• v.12 no.4
• /
• pp.142-146
• /
• 2023

We present a method for generating 3D structures and rendering objects by combining VAE (Variational Autoencoder) and GAN (Generative Adversarial Network). This approach focuses on generating and rendering 3D models with improved quality using residual learning as the learning method for the encoder. We deep stack the encoder layers to accurately reflect the features of the image and apply residual blocks to solve the problems of deep layers to improve the encoder performance. This solves the problems of gradient vanishing and exploding, which are problems when constructing a deep neural network, and creates a 3D model of improved quality. To accurately extract image features, we construct deep layers of the encoder model and apply the residual function to learning to model with more detailed information. The generated model has more detailed voxels for more accurate representation, is rendered by adding materials and lighting, and is finally converted into a mesh model. 3D models have excellent visual quality and accuracy, making them useful in various fields such as virtual reality, game development, and metaverse.

• KIPS Transactions on Software and Data Engineering
• /
• v.11 no.6
• /
• pp.237-244
• /
• 2022

In this paper, we present VAE-based recommendation using online behavior log and purchase history to overcome data sparsity and cold start. To generate a variable for customers' purchase history, embedding and dimensionality reduction are applied to the customers' purchase history. Also, Variational Autoencoders are applied to online behavior and purchase history. A total number of 12 variables are used, and nDCG is chosen for performance evaluation. Our experimental results showed that the proposed VAE-based recommendation outperforms SVD-based recommendation. Also, the generated purchase history variable improves the recommendation performance.

• The Journal of the Acoustical Society of Korea
• /
• v.43 no.2
• /
• pp.243-252
• /
• 2024

Among the Foley sound generation models that have recently begun to be studied, a sound generation technique using the Vector Quantized-Variational AutoEncoder (VQ-VAE) structure and generation model such as Pixelsnail are one of the important research subjects. On the other hand, in the field of deep learning-based acoustic signal compression, residual vector quantization technology is reported to be more suitable than the conventional VQ-VAE structure. Therefore, in this paper, we aim to study whether residual vector quantization technology can be effectively applied to the Foley sound generation. In order to tackle the problem, this paper applies the residual vector quantization technique to the conventional VQ-VAE-based Foley sound generation model, and in particular, derives a model that is compatible with the existing models such as Pixelsnail and does not increase computational resource consumption. In order to evaluate the model, an experiment was conducted using DCASE2023 Task7 data. The results show that the proposed model enhances about 0.3 of the Fréchet audio distance. Unfortunately, the performance enhancement was limited, which is believed to be due to the decrease in the resolution of time-frequency domains in order to do not increase consumption of the computational resources.

• Journal of the Korean Society of Systems Engineering
• /
• v.19 no.2
• /
• pp.145-156
• /
• 2023

Trajectory of ballistic missile is defined by inherent flight dynamics, which decided range and maneuvering characteristics. It is crucial to predict range and maneuvering characteristics of ballistic missile in KAMD (Korea Air and Missile Defense) to minimize damage due to ballistic missile attacks, Nowadays, needs for applying AI(Artificial Intelligence) technologies are increasing due to rapid developments of DNN(Deep Neural Networks) technologies. To apply these DNN technologies amount of data are required for superviesed learning, but trajectory data of ballistic missiles is limited because of security issues. Trajectory data could be considered as multivariate time series including many variables. And augmentation in time series data is a developing area of research. In this paper, we tried to augment trajectory data of ballistic missiles using recently developed methods. We used TimeVAE(Time Variational AutoEncoder) method and TimeGAN(Time Generative Adversarial Networks) to synthesize missile trajectory data. We also compare the results of two methods and analyse for future works.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.13 no.3
• /
• pp.197-205
• /
• 2020

Korean language has the characteristics that the pronunciation of phoneme units such as vowels and consonants are fixed and the pronunciation associated with a notation does not change, so that foreign learners can approach rather easily Korean language. However, when one pronounces words, phrases, or sentences, the pronunciation changes in a manner of a wide variation and complexity at the boundaries of syllables, and the association of notation and pronunciation does not hold any more. Consequently, it is very difficult for foreign learners to study Korean standard pronunciations. Despite these difficulties, it is believed that systematic analysis of pronunciation errors for Korean words is possible according to the advantageous observations that the relationship between Korean notations and pronunciations can be described as a set of firm rules without exceptions unlike other languages including English. In this paper, we propose a visualization framework which shows the differences between standard pronunciations and erratic ones as quantitative measures on the computer screen. Previous researches only show color representation and 3D graphics of speech properties, or an animated view of changing shapes of lips and mouth cavity. Moreover, the features used in the analysis are only point data such as the average of a speech range. In this study, we propose a method which can directly use the time-series data instead of using summary or distorted data. This was realized by using the deep learning-based technique which combines Self-organizing map, variational autoencoder model, and Markov model, and we achieved a superior performance enhancement compared to the method using the point-based data.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.9
• /
• pp.2927-2941
• /
• 2022

The primary task of machine fault diagnosis is to judge whether the current state is normal or damaged, so it is a typical binary classification problem with mutual exclusion. Mutually exclusive events and out-of-domain detection have one thing in common: there are two types of data and no intersection. We proposed a fusion model method to improve the accuracy of machine fault diagnosis, which is based on the mutual exclusivity of events and the commonality of out-of-distribution detection, and finally generalized to all binary classification problems. It is reported that the performance of a convolutional neural network (CNN) will decrease as the recognition type increases, so the variational auto-encoder (VAE) is used as the primary model. Two VAE models are used to train the machine's normal and fault sound data. Two reconstruction probabilities will be obtained during the test. The smaller value is transformed into a correction value of another value according to the mutually exclusive characteristics. Finally, the classification result is obtained according to the fusion algorithm. Filtering normal data features from fault data features is proposed, which shields the interference and makes the fault features more prominent. We confirm that good performance improvements have been achieved in the machine fault detection data set, and the results are better than most mainstream models.

• Information Systems Review
• /
• v.23 no.1
• /
• pp.23-43
• /
• 2021

Recently, many studies have been conducted to increase the accuracy of stock price prediction by analyzing candlestick charts using artificial intelligence techniques. However, these studies failed to consider the time-series characteristics of candlestick charts and to take into account the emotional state of market participants in data learning for stock price prediction. In order to overcome these limitations, this study produced input data by combining volatility index and candlestick charts to consider the emotional state of market participants, and used the data as input for a new method proposed on the basis of combining variantion autoencoder (VAE) and attention mechanisms for considering the time-series characteristics of candlestick chart. Fifty firms were randomly selected from the S&P 500 index and their stock prices were predicted to evaluate the performance of the method compared with existing ones such as convolutional neural network (CNN) or long-short term memory (LSTM). The results indicated the method proposed in this study showed superior performance compared to the existing ones. This study implied that the accuracy of stock price prediction could be improved by considering the emotional state of market participants and the time-series characteristics of the candlestick chart.

• Smart Media Journal
• /
• v.9 no.3
• /
• pp.80-89
• /
• 2020

Out - of - distribution (OOD) samples are frequently encountered when deploying a classification model in plenty of real-world machine learning-based applications. Those samples are normally sampling far away from the training distribution, but many classifiers still assign them high reliability to belong to one of the training categories. In this study, we address the problem of removing OOD examples by estimating marginal density estimation using variational autoencoder (VAE). We also investigate other proper methods, such as temperature scaling, Gaussian discrimination analysis, and label smoothing. We use Chonnam National University (CNU) weeds dataset as the in - distribution dataset and CIFAR-10, CalTeach as the OOD datasets. Quantitative results show that the proposed framework can reject the OOD test samples with a suitable threshold.

• Journal of Information Processing Systems
• /
• v.17 no.6
• /
• pp.1057-1070
• /
• 2021

Case-related news filtering is crucial in legal text mining and divides news into case-related and case-unrelated categories. Because case-related news originates from various fields and has different writing styles, it is difficult to establish complete filtering rules or keywords for data collection. In addition, the labeled corpus for case-related news is sparse; therefore, to train a high-performance classification model, it is necessary to annotate the corpus. To address this challenge, we propose topic-enhanced positive-unlabeled learning, which selects positive and negative samples guided by topics. Specifically, a topic model based on a variational autoencoder (VAE) is trained to extract topics from unlabeled samples. By using these topics in the iterative process of positive-unlabeled (PU) learning, the accuracy of identifying case-related news can be improved. From the experimental results, it can be observed that the F1 value of our method on the test set is 1.8% higher than that of the PU learning baseline model. In addition, our method is more robust with low initial samples and high iterations, and compared with advanced PU learning baselines such as nnPU and I-PU, we obtain a 1.1% higher F1 value, which indicates that our method can effectively identify case-related news.