• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.527-529
• /
• 2018

초기 자연언어처리에 FNN(Feedforward Neural Network)을 적용한 연구들에 비해 LSTM(Long Short-Term Memory)은 현재 시점의 정보뿐만 아니라 이전 시점의 정보를 담고 있어 문장을 이루는 어절들, 어절을 이루는 형태소 등 순차적인(sequential) 데이터를 처리하는데 좋은 성능을 보인다. 본 논문에서는 스택과 버퍼에 있는 어절을 양방향 LSTM encoding을 이용한 representation으로 표현하여 전이기반 의존구문분석에 적용하여 현재 UAS 89.4%의 정확도를 보였고, 자질 추가 및 정제작업을 통해 성능이 개선될 것으로 보인다.

• IE interfaces
• /
• v.10 no.3
• /
• pp.125-132
• /
• 1997

We proposed a 2-step neural network approach for estimating machining parameters of electric discharge wire cut. The first step net, which is described as a backward neuro-estimation, is designed for estimating coarse cutting parameters while the second phase net, as a polishing forward neuro-estimation, is utilized for determining fine parameters. Sequential estimation procedure, based on backward and forward net, is performed using the net's approximation capability which is M to 1 and 1 to M mapping property. Experimental results an given to evaluate the accuracy of the proposed 2-step neuro-estimation.

• International Journal of Fuzzy Logic and Intelligent Systems
• /
• v.6 no.4
• /
• pp.271-276
• /
• 2006

To implement a generalization of value functions in Adaptive Search Element (ASE)-reinforcement learning, CMAC (Cerebellar Model Articulation Controller) is integrated into ASE controller. ASE-reinforcement learning scheme is briefly studied to discuss how CMAC is integrated into ASE controller. Neighbourhood Sequential Training for CMAC is utilized to establish the look-up table and to produce discrete control outputs. In computer simulation, an ASE controller and a couple of ASE-CMAC neural network are trained to balance the inverted pendulum on a cart. The number of trials until the controllers are established and the learning performance of the controllers are evaluated to find that generalization ability of the CMAC improves the speed of the ASE-reinforcement learning enough to realize the cartpole control system.

• Journal of the Semiconductor & Display Technology
• /
• v.20 no.1
• /
• pp.83-87
• /
• 2021

By the increase of internet usage, communicating online became an everyday thing. Thereby various people have experienced profanity by anonymous users. Nowadays lots of studies tried to solve this problem using artificial intelligence, but most of the solutions were for non-real time situations. In this paper, we propose a Telegram plugin that detects swear words using word2vec, and an algorithm to find the target of the sentence. We vectorized the input sentence to find connections with other similar words, then inputted the value to the pre-trained CNN (Convolutional Neural Network) model to detect any swears. For target recognition we proposed a sequential algorithm based on KoNLPY.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.4
• /
• pp.633-636
• /
• 2022

The reduction of energy consumption at the base station (BS) has become more important recently. In this paper, we consider the adaptive muting of the antennas based on the predicted future traffic load to reduce the energy consumption where the number of active antennas is adaptively adjusted according to the predicted future traffic load. Given that traffic load is sequential data, three different RNN structures, namely long-short term memory (LSTM), gated recurrent unit (GRU), and bidirectional LSTM (Bi-LSTM) are considered for the future traffic load prediction. Through the performance evaluation based on the actual traffic load collected from the Afghanistan telecom company, we confirm that the traffic load can be estimated accurately and the overall power consumption can also be reduced significantly using the antenna musing.

• ETRI Journal
• /
• v.45 no.6
• /
• pp.929-938
• /
• 2023

The detection of all the symbols transmitted simultaneously in multiuser systems using limited wireless resources is challenging. Traditional model-based methods show high performance with perfect channel state information (CSI); however, severe performance degradation will occur if perfect CSI cannot be acquired. In contrast, data-driven methods perform slightly worse than model-based methods in terms of symbol error ratio performance in perfect CSI states; however, they are also able to overcome extreme performance degradation in imperfect CSI states. This study proposes a novel deep learning-based method by improving a state-of-the-art data-driven technique called deep soft interference cancellation (DSIC). The enhanced DSIC (EDSIC) method detects multiuser symbols in a fully sequential manner and uses an efficient neural network structure to ensure high performance. Additionally, error-propagation mitigation techniques are used to ensure robustness against channel uncertainty. The EDSIC guarantees a performance that is very close to the optimal performance of the existing model-based methods in perfect CSI environments and the best performance in imperfect CSI environments.

• ETRI Journal
• /
• v.46 no.4
• /
• pp.648-659
• /
• 2024

Author verification is challenging because of the diversity in writing styles. We propose an enhanced handwriting verification method that combines handcrafted and automatically extracted features. The method uses a genetic algorithm to reduce the dimensionality of the feature set. We consider offline Bangla handwriting content and evaluate the proposed method using handcrafted features with a simple logistic regression, radial basis function network, and sequential minimal optimization as well as automatically extracted features using a convolutional neural network. The handcrafted features outperform the automatically extracted ones, achieving an average verification accuracy of 94.54% for 100 writers. The handcrafted features include Radon transform, histogram of oriented gradients, local phase quantization, and local binary patterns from interwriter and intrawriter content. The genetic algorithm reduces the feature dimensionality and selects salient features using a support vector machine. The top five experimental results are obtained from the optimal feature set selected using a consensus strategy. Comparisons with other methods and features confirm the satisfactory results.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.30B no.9
• /
• pp.94-105
• /
• 1993

In this paper propose a Koaren character recognition system using a neural network is proposed. This system is a multilayer neural network based on the masking field model which consists of a input layer, four feature extraction layers which extracts type, direction, stroke, and connection features, and an output layer which gives us recognized character codes. First, 4x4 subpatterns of an NxN character pattern stored in the input buffer are applied into the feature extraction layers sequentially. Then, each of feature extraction layers extracts sequentially features such as type, direction, stroke, and connection, respectively. Type features for direction and connection are extracted by the type feature extraction layer, direction features for stroke by the direction feature extraction layer and stroke and connection features for stroke by the direction feature extraction layer and stroke and connection features for the recongnition of character by the stroke and the connection feature extractions layers, respectively. The stroke and connection features are saved in the sequential buffer layer sequentially and using these features the characters are recognized in the output layer. The recognition results of this system by tests with 8 single consonants and 6 single vowels are promising.

• Korean Journal of Metals and Materials
• /
• v.56 no.11
• /
• pp.813-821
• /
• 2018

The basic oxygen furnace (BOF) steelmaking process in the steel industry is highly complicated, and subject to variations in raw material composition. During the BOF steelmaking process, it is essential to maintain the carbon content and the endpoint temperature at their set points in the liquid steel. This paper presents intelligent models used to estimate the endpoint temperature in the basic oxygen furnace (BOF) steelmaking process. An artificial neural network (ANN) model and a least-squares support vector machine (LSSVM) model are proposed and their estimation performance compared. The classical partial least-squares (PLS) method was also compared with the others. Results of the estimations using the ANN, LSSVM and PLS models were compared with the operation data, and the root-mean square error (RMSE) for each model was calculated to evaluate estimation performance. The RMSE of the LSSVM model 15.91, which turned out to be the best estimation. RMSE values for the ANN and PLS models were 17.24 and 21.31, respectively, indicating their relative estimation performance. The essential input parameters used in the models can be selected by sensitivity analysis. The RMSE for each model was calculated again after a sequential input selection process was used to remove insignificant input parameters. The RMSE of the LSSVM was then 13.21, which is better than the previous RMSE with all 16 parameters. The results show that LSSVM model using 13 input parameters can be utilized to calculate the required values for oxygen volume and coolant needed to optimally adjust the steel target temperature.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2021.06a
• /
• pp.131-131
• /
• 2021

Deep learning models, especially those based on long short-term memory (LSTM), have presented their superiority in addressing time series data issues recently. This study aims to comprehensively evaluate the performance of deep learning models that belong to the supervised learning category in streamflow prediction. Therefore, six deep learning models-standard LSTM, standard gated recurrent unit (GRU), stacked LSTM, bidirectional LSTM (BiLSTM), feed-forward neural network (FFNN), and convolutional neural network (CNN) models-were of interest in this study. The Red River system, one of the largest river basins in Vietnam, was adopted as a case study. In addition, deep learning models were designed to forecast flowrate for one- and two-day ahead at Son Tay hydrological station on the Red River using a series of observed flowrate data at seven hydrological stations on three major river branches of the Red River system-Thao River, Da River, and Lo River-as the input data for training, validation, and testing. The comparison results have indicated that the four LSTM-based models exhibit significantly better performance and maintain stability than the FFNN and CNN models. Moreover, LSTM-based models may reach impressive predictions even in the presence of upstream reservoirs and dams. In the case of the stacked LSTM and BiLSTM models, the complexity of these models is not accompanied by performance improvement because their respective performance is not higher than the two standard models (LSTM and GRU). As a result, we realized that in the context of hydrological forecasting problems, simple architectural models such as LSTM and GRU (with one hidden layer) are sufficient to produce highly reliable forecasts while minimizing computation time because of the sequential data nature.