• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.12
• /
• pp.1769-1776
• /
• 2022

Automatically summarizing long sentences is an important technique. The BART model is one of the widely used models in the summarization task. In general, in order to generate a summarization model of a specific domain, fine-tuning is performed by re-training a language model trained on a large dataset to fit the domain. The fine-tuning is usually done by changing the number of nodes in the last fully connected layer. However, in this paper, we propose a fine-tuning method by adding an attention layer, which has been recently applied to various models and shows good performance. In order to evaluate the performance of the proposed method, various experiments were conducted, such as accumulating layers deeper, fine-tuning without skip connections during the fine tuning process, and so on. As a result, the BART model using two attention layers with skip connection shows the best score.

• The Bulletin of The Korean Astronomical Society
• /
• v.46 no.2
• /
• pp.46.3-47
• /
• 2021

Recently a multilayer spectral inversion (MLSI) model has been proposed to infer the physical parameters of plasmas in the solar chromosphere. The inversion solves a three-layer radiative transfer model using the strong absorption line profiles, H alpha and Ca II 8542 Å, taken by the Fast Imaging Solar Spectrograph (FISS). The model successfully provides the physical plasma parameters, such as source functions, Doppler velocities, and Doppler widths in the layers of the photosphere to the chromosphere. However, it is quite expensive to apply the MLSI to a huge number of line profiles. For example, the calculating time is an hour to several hours depending on the size of the scan raster. We apply deep neural network (DNN) to the inversion code to reduce the cost of calculating the physical parameters. We train the models using pairs of absorption line profiles from FISS and their 13 physical parameters (source functions, Doppler velocities, Doppler widths in the chromosphere, and the pre-determined parameters for the photosphere) calculated from the spectral inversion code for 49 scan rasters (~2,000,000 dataset) including quiet and active regions. We use fully connected dense layers for training the model. In addition, we utilize a skip connection to avoid a problem of vanishing gradients. We evaluate the model by comparing the pairs of absorption line profiles and their inverted physical parameters from other quiet and active regions. Our result shows that the deep learning model successfully reproduces physical parameter maps of a scan raster observation per second within 15% of mean absolute percentage error and the mean squared error of 0.3 to 0.003 depending on the parameters. Taking this advantage of high performance of the deep learning model, we plan to provide the physical parameter maps from the FISS observations to understand the chromospheric plasma conditions in various solar features.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.6
• /
• pp.703-709
• /
• 2019

This paper proposes speech recognition systems employing Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) structures combined with Hidden Markov Moldel (HMM) to effectively recognize the speech of VeloPharyngeal Insufficiency (VPI) patients, and compares the recognition performance of the systems to the Gaussian Mixture Model (GMM-HMM) and fully-connected Deep Neural Network (DNNHMM) based speech recognition systems. In this paper, the initial model is trained using normal speakers' speech and simulated VPI speech is used for generating a prior model for speaker adaptation. For VPI speaker adaptation, selected layers are trained in the CNN-HMM based model, and dropout regulatory technique is applied in the LSTM-HMM based model, showing 3.68 % improvement in recognition accuracy. The experimental results demonstrate that the proposed LSTM-HMM-based speech recognition system is effective for VPI speech with small-sized speech data, compared to conventional GMM-HMM and fully-connected DNN-HMM system.

• The Transactions of the Korea Information Processing Society
• /
• v.13 no.4
• /
• pp.199-207
• /
• 2024

Local reservoirs are crucial sources for agricultural water supply, necessitating stable water level management to prepare for extreme climate conditions such as droughts. Water level prediction is significantly influenced by local climate characteristics, such as localized rainfall, as well as seasonal factors including cropping times, making it essential to understand the correlation between input and output data as much as selecting an appropriate prediction model. In this study, extensive multivariate data from over 400 reservoirs in Jeollabuk-do from 1991 to 2022 was utilized to train and validate a water level prediction model that comprehensively reflects the complex hydrological and climatological environmental factors of each reservoir, and to analyze the impact of each input feature on the prediction performance of water levels. Instead of focusing on improvements in water level performance through neural network structures, the study adopts a basic Feedforward Neural Network composed of fully connected layers, batch normalization, dropout, and activation functions, focusing on the correlation between multivariate input data and prediction performance. Additionally, most existing studies only present short-term prediction performance on a daily basis, which is not suitable for practical environments that require medium to long-term predictions, such as 10 days or a month. Therefore, this study measured the water level prediction performance up to one month ahead through a recursive method that uses daily prediction values as the next input. The experiment identified performance changes according to the prediction period and analyzed the impact of each input feature on the overall performance based on an Ablation study.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.6
• /
• pp.592-599
• /
• 2020

This paper proposes a Convolutional Recurrent Neural Net (CRNN) structure that can simultaneously reflect both static and dynamic characteristics of seismic waveforms for various earthquake events classification. Addressing various earthquake events, including not only micro-earthquakes and artificial-earthquakes but also macro-earthquakes, requires both effective feature extraction and a classifier that can discriminate seismic waveform under noisy environment. First, we extract the static characteristics of seismic waveform through an attention-based convolution layer. Then, the extracted feature-map is sequentially injected as input to a multi-input single-output Long Short-Term Memory (LSTM) network structure to extract the dynamic characteristic for various seismic event classifications. Subsequently, we perform earthquake events classification through two fully connected layers and softmax function. Representative experimental results using domestic and foreign earthquake database show that the proposed model provides an effective structure for various earthquake events classification.

• Journal of the Korean Society of Manufacturing Process Engineers
• /
• v.19 no.7
• /
• pp.7-15
• /
• 2020

Recently, the fourth industrial revolution has been presented as a new paradigm and additive manufacturing (AM) has become one of the most important topics. For this reason, process monitoring for each cross-sectional layer of additive metal manufacturing is important. Particularly, deep learning can train a machine to analyze, optimize, and repair defects. In this paper, image classification is proposed by learning images of defects in the metal cross sections using the convolution neural network (CNN) image labeling algorithm. Defects were classified into three categories: crack, porosity, and hole. To overcome a lack-of-data problem, the amount of learning data was augmented using a data augmentation algorithm. This augmentation algorithm can transform an image to 180 images, increasing the learning accuracy. The number of training and validation images was 25,920 (80 %) and 6,480 (20 %), respectively. An optimized case with a combination of fully connected layers, an optimizer, and a loss function, showed that the model accuracy was 99.7 % and had a success rate of 97.8 % for 180 test images. In conclusion, image labeling was successfully performed and it is expected to be applied to automated AM process inspection and repair systems in the future.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.5
• /
• pp.1431-1445
• /
• 2022

We construct a deep neural network model named ECGResNet. This model can diagnosis diseases based on 12-lead ECG data of eight common cardiovascular diseases with a high accuracy. We chose the 16 Blocks of ResNet50 as the main body of the model and added the Squeeze-and-Excitation module to learn the data information between channels adaptively. We modified the first convolutional layer of ResNet50 which has a convolutional kernel of 7 to a superposition of convolutional kernels of 8 and 16 as our feature extraction method. This way allows the model to focus on the overall trend of the ECG signal while also noticing subtle changes. The model further improves the accuracy of cardiovascular and cerebrovascular disease classification by using a fully connected layer that integrates factors such as gender and age. The ECGResNet model adds Dropout layers to both the residual block and SE module of ResNet50, further avoiding the phenomenon of model overfitting. The model was eventually trained using a five-fold cross-validation and Flooding training method, with an accuracy of 95% on the test set and an F1-score of 0.841.We design a new deep neural network, innovate a multi-scale feature extraction method, and apply the SE module to extract features of ECG data.

• Journal of the Korean Association of Geographic Information Studies
• /
• v.2 no.2
• /
• pp.69-78
• /
• 1999

In this paper, a neural network approach to forecast Korean regional precipitation is presented. We first analyze the characteristics of the conventional models for time series prediction, and then propose a new model and its learning method for the precipitation forecast. The proposed model is a layered network in which the outputs of a layer are buffered within a given period time and then fed fully connected to the upper layer. This study adopted the dual connections between two layers for the model. The network behavior and learning algorithm for the model are also described. The dual connection structure plays the role of the bias of the ordinary Multi-Layer Perceptron(MLP), and reflects the relationships among the features effectively. From these advantageous features, the model provides the learning efficiency in comparison with the FIR network, which is the most popular model for time series prediction. We have applied the model to the monthly and seasonal forecast of precipitation. The precipitation data and SST(Sea Surface Temperature) data for several decades are used as the learning pattern for the neural network predictor. The experimental results have shown the validity of the proposed model.

• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.10
• /
• pp.373-380
• /
• 2022

This paper presents a deep learning based group synchronization that supports networked immersive interactions between remote users. The goal of group synchronization is to enable all participants to synchronously interact with others for increasing user presence Most previous methods focus on NTP-based clock synchronization to enhance time accuracy. Moving average filters are used to control media playout time on the synchronization server. As an example, the exponentially weighted moving average(EWMA) would be able to track and estimate accurate playout time if the changes in input data are not significant. However it needs more time to be stable for any given change over time due to codec and system loads or fluctuations in network status. To tackle this problem, this work proposes the Deep Group Synchronization(DeepGroupSync), a group synchronization based on deep learning that models important features from the data. This model consists of two Gated Recurrent Unit(GRU) layers and one fully-connected layer, which predicts an optimal playout time by utilizing the sequential playout delays. The experiments are conducted with an existing method that uses the EWMA and the proposed method that uses the DeepGroupSync. The results show that the proposed method are more robust against unpredictable or rapid network condition changes than the existing method.

• Korean Journal of Remote Sensing
• /
• v.39 no.2
• /
• pp.207-221
• /
• 2023

This study suggests deep neural network models for estimating air temperature with Level 1B (L1B) datasets of GEO-KOMPSAT-2A (GK-2A). The temperature at 1.5 m above the ground impact not only daily life but also weather warnings such as cold and heat waves. There are many studies to assume the air temperature from the land surface temperature (LST) retrieved from satellites because the air temperature has a strong relationship with the LST. However, an algorithm of the LST, Level 2 output of GK-2A, works only clear sky pixels. To overcome the cloud effects, we apply a deep neural network (DNN) model to assume the air temperature with L1B calibrated for radiometric and geometrics from raw satellite data and compare the model with a linear regression model between LST and air temperature. The root mean square errors (RMSE) of the air temperature for model outputs are used to evaluate the model. The number of 95 in-situ air temperature data was 2,496,634 and the ratio of datasets paired with LST and L1B show 42.1% and 98.4%. The training years are 2020 and 2021 and 2022 is used to validate. The DNN model is designed with an input layer taking 16 channels and four hidden fully connected layers to assume an air temperature. As a result of the model using 16 bands of L1B, the DNN with RMSE 2.22℃ showed great performance than the baseline model with RMSE 3.55℃ on clear sky conditions and the total RMSE including overcast samples was 3.33℃. It is suggested that the DNN is able to overcome cloud effects. However, it showed different characteristics in seasonal and hourly analysis and needed to append solar information as inputs to make a general DNN model because the summer and winter seasons showed a low coefficient of determinations with high standard deviations.