• The Journal of the Convergence on Culture Technology
• /
• v.7 no.3
• /
• pp.591-598
• /
• 2021

It's possible to get the word-vector by the statistical SVD or deep-learning CBOW and LSTM methods and theses ones learn the contexts of forward/backward words or the sequence of following words. It's used to analyze the poems by Ki Hyung-do with similar words recommended by the word-vector showing the core images of the poetry. It seems at first sight that the words don't go well with the images but they express the similar style described by the reference words once you look close the contexts of the specific poems. The word-vector can analogize the words having the same relations with the ones between the representative words for the core images of the poems. Therefore you can analyze the poems in depth and in variety with the similarity and analogy operations by the word-vector estimated with the statistical SVD or deep-learning CBOW and LSTM methods.

• Korean Journal of Agricultural Science
• /
• v.46 no.1
• /
• pp.67-78
• /
• 2019

The purpose of this study was to predict the water quality using the RNN (recurrent neutral network) and LSTM (long short-term memory). These are advanced forms of machine learning algorithms that are better suited for time series learning compared to artificial neural networks; however, they have not been investigated before for water quality prediction. Three water quality indexes, the BOD (biochemical oxygen demand), COD (chemical oxygen demand), and SS (suspended solids) are predicted by the RNN and LSTM. TensorFlow, an open source library developed by Google, was used to implement the machine learning algorithm. The Okcheon observation point in the Geum River basin in the Republic of Korea was selected as the target point for the prediction of the water quality. Ten years of daily observed meteorological (daily temperature and daily wind speed) and hydrological (water level and flow discharge) data were used as the inputs, and irregularly observed water quality (BOD, COD, and SS) data were used as the learning materials. The irregularly observed water quality data were converted into daily data with the linear interpolation method. The water quality after one day was predicted by the machine learning algorithm, and it was found that a water quality prediction is possible with high accuracy compared to existing physical modeling results in the prediction of the BOD, COD, and SS, which are very non-linear. The sequence length and iteration were changed to compare the performances of the algorithms.

• Journal of the Korea Convergence Society
• /
• v.8 no.12
• /
• pp.55-62
• /
• 2017

The Named Entity Recognition system is a system that recognizes words or phrases with object names such as personal name (PS), place name (LC), and group name (OG) in the document as corresponding object names. Traditional approaches to named entity recognition include statistical-based models that learn models based on hand-crafted features. Recently, it has been proposed to construct the qualities expressing the sentence using models such as deep-learning based Recurrent Neural Networks (RNN) and long-short term memory (LSTM) to solve the problem of sequence labeling. In this research, to improve the performance of the Korean named entity recognition system, we used a hand-crafted feature, part-of-speech tagging information, and pre-built lexicon information to augment features for representing sentence. Experimental results show that the proposed method improves the performance of Korean named entity recognition system. The results of this study are presented through github for future collaborative research with researchers studying Korean Natural Language Processing (NLP) and named entity recognition system.

• Journal of Korean Society of Disaster and Security
• /
• v.14 no.4
• /
• pp.17-27
• /
• 2021

Streamflow prediction is a very vital disaster mitigation approach for effective flood management and water resources planning. Lately, torrential rainfall caused by climate change has been reported to have increased globally, thereby causing enormous infrastructural loss, properties and lives. This study evaluates the contribution of rainfall to streamflow prediction in normal and peak rainfall scenarios, typical of the recent flood at Piney Resort in Vernon, Hickman County, Tennessee, United States. Daily streamflow, water level, and rainfall data for 20 years (2000-2019) from two USGS gage stations (03602500 upstream and 03599500 downstream) of the Piney River watershed were obtained, preprocesssed and fitted with Long short term memory (LSTM) model. Tensorflow and Keras machine learning frameworks were used with Python to predict streamflow values with a sequence size of 14 days, to determine whether the model could have predicted the flooding event in August 21, 2021. Model skill analysis showed that LSTM model with full data (water level, streamflow and rainfall) performed better than the Naive Model except some rainfall models, indicating that only rainfall is insufficient for streamflow prediction. The final LSTM model recorded optimal NSE and RMSE values of 0.68 and 13.84 m³/s and predicted peak flow with the lowest prediction error of 11.6%, indicating that the final model could have predicted the flood on August 24, 2021 given a peak rainfall scenario. Adequate knowledge of rainfall patterns will guide hydrologists and disaster prevention managers in designing efficient early warning systems and policies aimed at mitigating flood risks.

• Journal of the Korean Society of Systems Engineering
• /
• v.18 no.2
• /
• pp.75-93
• /
• 2022

Machine learning (ML) data-driven meta-model is proposed as a surrogate model to reduce the excessive computational cost of the physics-based model and facilitate the real-time prediction of a nuclear power plant's transient response. To forecast the transient response three machine learning (ML) meta-models based on recurrent neural networks (RNNs); specifically, Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), and a sequence combination of Convolutional Neural Network (CNN) and LSTM are developed. The chosen accident scenario is a control element assembly withdrawal at power concurrent with the Loss Of Offsite Power (LOOP). The transient response was obtained using the best estimate thermal hydraulics code, MARS-KS, and cross-validated against the Design and control document (DCD). DAKOTA software is loosely coupled with MARS-KS code via a python interface to perform the Best Estimate Plus Uncertainty Quantification (BEPU) analysis and generate a time series database of the system response to train, test and validate the ML meta-models. Key uncertain parameters identified as required by the CASU methodology were propagated using the non-parametric Monte-Carlo (MC) random propagation and Latin Hypercube Sampling technique until a statistically significant database (181 samples) as required by Wilk's fifth order is achieved with 95% probability and 95% confidence level. The three ML RNN models were built and optimized with the help of the Talos tool and demonstrated excellent performance in forecasting the most probable NPP transient response. This research was guided by the Systems Engineering (SE) approach for the systematic and efficient planning and execution of the research.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2021.06a
• /
• pp.23-25
• /
• 2021

This paper describes an approach of composing music with multiple AI composers. This approach enriches more the creativity space of artificial intelligence music composition than using only one composer. This paper presents a simple example with 2 different deep learning composers working together for composing one music. For the experiment, the two composers adopt the same deep learning architecture of an LSTM model trained with different data. The output of a composer is a sequence of notes. Each composer alternatively appends its output to the resulting music which is input to both the composers. Experiments compare different music generated by the proposed multiple composer approach with the traditional one composer approach.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.7
• /
• pp.1759-1772
• /
• 2023

Chinese named entity recognition (NER) is a challenging work that seeks to find, recognize and classify various types of information elements in unstructured text. Due to the Chinese text has no natural boundary like the spaces in the English text, Chinese named entity identification is much more difficult. At present, most deep learning based NER models are developed using a bidirectional long short-term memory network (BiLSTM), yet the performance still has some space to improve. To further improve their performance in Chinese NER tasks, we propose a new NER model, IDCNN-BiLSTM-Highway, which is a combination of the BiLSTM, the iterated dilated convolutional neural network (IDCNN) and the highway network. In our model, IDCNN is used to achieve multiscale context aggregation from a long sequence of words. Highway network is used to effectively connect different layers of networks, allowing information to pass through network layers smoothly without attenuation. Finally, the global optimum tag result is obtained by introducing conditional random field (CRF). The experimental results show that compared with other popular deep learning-based NER models, our model shows superior performance on two Chinese NER data sets: Resume and Yidu-S4k, The F1-scores are 94.98 and 77.59, respectively.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.14 no.5
• /
• pp.277-286
• /
• 2019

In this paper, we propose an efficient hyperplane generation technique to classify human activity from combination of events and sequence information obtained from multiple-event sensors. By generating hyperplane efficiently, our machine learning algorithm classify with less memory and run time than the LSVM (Linear Support Vector Machine) for embedded system. Because the fact that light weight and high speed algorithm is one of the most critical issue in the IoT, the study can be applied to smart home to predict human activity and provide related services. Our approach is based on reducing numbers of hyperplanes and utilizing robust string comparing algorithm. The proposed method results in reduction of memory consumption compared to the conventional ML (Machine Learning) algorithms; 252 times to LSVM and 34,033 times to LSTM (Long Short-Term Memory), although accuracy is decreased slightly. Thus our method showed outstanding performance on accuracy per hyperplane; 240 times to LSVM and 30,520 times to LSTM. The binarized image is then divided into groups, where each groups are converted to binary number, in order to reduce the number of comparison done in runtime process. The binary numbers are then converted to string. The test data is evaluated by converting to string and measuring similarity between hyperplanes using Levenshtein algorithm, which is a robust dynamic string comparing algorithm. This technique reduces runtime and enables the proposed algorithm to become 27% faster than LSVM, and 90% faster than LSTM.

• International Journal of Contents
• /
• v.15 no.4
• /
• pp.8-15
• /
• 2019

This paper presents a solution of the 'Quick, Draw! Doodle Recognition Challenge' hosted by Google. Doodles are drawings comprised of concrete representational meaning or abstract lines creatively expressed by individuals. In this challenge, a doodle is presented as a sequence of sketches. From the view of at the sketch level, to learn the pattern of strokes representing a doodle, we propose a sequential model stacked with multiple convolution layers and Long Short-Term Memory (LSTM) cells following the attention mechanism [15]. From the view at the image level, we use multiple models pre-trained on ImageNet to recognize the doodle. Finally, an ensemble and a post-processing method using the minimum cost flow algorithm are introduced to combine multiple models in achieving better results. In this challenge, our solutions garnered 11th place among 1,316 teams. Our performance was 0.95037 MAP@3, only 0.4% lower than the winner. It demonstrates that our method is very competitive. The source code for this competition is published at: https://github.com/ngxbac/Kaggle-QuickDraw.