• Journal of Intelligence and Information Systems
• /
• v.22 no.2
• /
• pp.127-142
• /
• 2016

Deep learning model is a kind of neural networks that allows multiple hidden layers. There are various deep learning architectures such as convolutional neural networks, deep belief networks and recurrent neural networks. Those have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. Among those architectures, convolutional neural networks and recurrent neural networks are classified as the supervised learning model. And in recent years, those supervised learning models have gained more popularity than unsupervised learning models such as deep belief networks, because supervised learning models have shown fashionable applications in such fields mentioned above. Deep learning models can be trained with backpropagation algorithm. Backpropagation is an abbreviation for "backward propagation of errors" and a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of an error function with respect to all the weights in the network. The gradient is fed to the optimization method which in turn uses it to update the weights, in an attempt to minimize the error function. Convolutional neural networks use a special architecture which is particularly well-adapted to classify images. Using this architecture makes convolutional networks fast to train. This, in turn, helps us train deep, muti-layer networks, which are very good at classifying images. These days, deep convolutional networks are used in most neural networks for image recognition. Convolutional neural networks use three basic ideas: local receptive fields, shared weights, and pooling. By local receptive fields, we mean that each neuron in the first(or any) hidden layer will be connected to a small region of the input(or previous layer's) neurons. Shared weights mean that we're going to use the same weights and bias for each of the local receptive field. This means that all the neurons in the hidden layer detect exactly the same feature, just at different locations in the input image. In addition to the convolutional layers just described, convolutional neural networks also contain pooling layers. Pooling layers are usually used immediately after convolutional layers. What the pooling layers do is to simplify the information in the output from the convolutional layer. Recent convolutional network architectures have 10 to 20 hidden layers and billions of connections between units. Training deep learning networks has taken weeks several years ago, but thanks to progress in GPU and algorithm enhancement, training time has reduced to several hours. Neural networks with time-varying behavior are known as recurrent neural networks or RNNs. A recurrent neural network is a class of artificial neural network where connections between units form a directed cycle. This creates an internal state of the network which allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequences of inputs. Early RNN models turned out to be very difficult to train, harder even than deep feedforward networks. The reason is the unstable gradient problem such as vanishing gradient and exploding gradient. The gradient can get smaller and smaller as it is propagated back through layers. This makes learning in early layers extremely slow. The problem actually gets worse in RNNs, since gradients aren't just propagated backward through layers, they're propagated backward through time. If the network runs for a long time, that can make the gradient extremely unstable and hard to learn from. It has been possible to incorporate an idea known as long short-term memory units (LSTMs) into RNNs. LSTMs make it much easier to get good results when training RNNs, and many recent papers make use of LSTMs or related ideas.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.05a
• /
• pp.401-403
• /
• 2019

In this paper, we propose a generative adversarial networks (GAN) based text-to-image generating method. In many natural language processing tasks, which word expressions are determined by their term frequency -inverse document frequency scores. Word2Vec is a type of neural network model that, in the case of an unlabeled corpus, produces a vector that expresses semantics for words in the corpus and an image is generated by GAN training according to the obtained vector. Thanks to the understanding of the word we can generate higher and more realistic images. Our GAN structure is based on deep convolution neural networks and pixel recurrent neural networks. Comparing the generated image with the real image, we get about 88% similarity on the Oxford-102 flowers dataset.

• Annual Conference on Human and Language Technology
• /
• 2016.10a
• /
• pp.67-71
• /
• 2016

본 논문에서는 word2vec과 doc2vec을 함께 CNN에 적용한 문서 분류 방안을 제안한다. 먼저 어절, 형태소, WPM(Word Piece Model)을 각각 사용하여 생성한 토큰(token)으로 doc2vec을 활용하여 문서를 vector로 표현한 후, 초보적인 문서 분류에 적용한 결과 WPM이 분류율 79.5%가 되어 3가지 방법 중 최고 성능을 보였다. 다음으로 CNN의 입력자질로써 WPM을 이용하여 생성한 토큰을 활용한 word2vec을 범주 10개의 문서 분류에 사용한 실험과 doc2vec을 함께 사용한 실험을 수행하였다. 실험 결과 word2vec만을 활용하였을 때 86.89%의 분류율을 얻었고, doc2vec을 함께 적용한 결과 89.51%의 분류율을 얻었다. 따라서 제안한 모델을 통해서 분류율이 2.62% 향상됨을 확인하였다.

• 한국어정보학회:학술대회논문집
• /
• 2016.10a
• /
• pp.217-219
• /
• 2016

본 논문은 형태소 분석 및 품사 태깅을 위해 seq2seq 주의집중 모델을 이용하는 접근 방법에 대하여 기술한다. seq2seq 모델은 인코더와 디코더로 분할되어 있고, 일반적으로 RNN(recurrent neural network)를 기반으로 한다. 형태소 분석 및 품사 태깅을 위해 seq2seq 모델의 학습 단계에서 음절 시퀀스는 인코더의 입력으로, 각 음절에 해당하는 품사 태깅 시퀀스는 디코더의 출력으로 사용된다. 여기서 음절 시퀀스와 품사 태깅 시퀀스의 대응관계는 주의집중(attention) 모델을 통해 접근하게 된다. 본 연구는 사전 정보나 자질 정보와 같은 추가적 리소스를 배제한 end-to-end 접근 방법의 실험 결과를 제시한다. 또한, 디코딩 단계에서 빔(beam) 서치와 같은 추가적 프로세스를 배제하는 접근 방법을 취한다.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.488-495
• /
• 2021

We propose a speech recognition system based on conformer. Conformer is known to be convolution-augmented transformer, which combines transfer model for capturing global information with Convolution Neural Network (CNN) for exploiting local feature effectively. The baseline system is developed to be a transfer-based speech recognition using Long Short-Term Memory (LSTM)-based language model. The proposed system is a system which uses conformer instead of transformer with transformer-based language model. When Electronics and Telecommunications Research Institute (ETRI) speech corpus in AI-Hub is used for our evaluation, the proposed system yields 5.7 % of Character Error Rate (CER) while the baseline system results in 11.8 % of CER. Even though speech corpus is extended into other domain of AI-hub such as NHNdiguest speech corpus, the proposed system makes a robust performance for two domains. Throughout those experiments, we can prove a validation of the proposed system.

• 한국어정보학회:학술대회논문집
• /
• 2016.10a
• /
• pp.263-267
• /
• 2016

채팅 시스템은 사람이 사용하는 언어로 컴퓨터와 의사소통을 하는 시스템이다. 최근 딥 러닝이 큰 화두가 되면서 다양한 채팅 시스템에 관한 연구가 빠르게 진행 되고 있다. 본 논문에서는 문장을 Recurrent Neural Network기반 의사형태소 분석기로 분리하고 Attention mechanism Encoder-Decoder Model의 입력으로 사용하는 채팅 시스템을 제안한다. 채팅 데이터를 통한 실험에서 사용자 문장이 짧은 경우는 답변이 잘 나오는 것을 확인하였으나 긴 문장에 대해서는 문법에 맞지 않는 문장이 생성되는 것을 알 수 있었다.

• 제어로봇시스템학회:학술대회논문집
• /
• 2003.10a
• /
• pp.1115-1120
• /
• 2003

The present paper shows the possible development of particle swarm optimization (PSO) based fuzzy-neural networks (FNN) which can be employed as an important building block in real life robot systems, controlled by voice-based commands. The PSO is employed to train the FNNs which can accurately output the crisp control signals for the robot systems, based on fuzzy linguistic spoken language commands, issued by an user. The FNN is also trained to capture the user spoken directive in the context of the present performance of the robot system. Hidden Markov Model (HMM) based automatic speech recognizers are developed, as part of the entire system, so that the system can identify important user directives from the running utterances. The system is successfully employed in a real life situation for motion control of a redundant manipulator.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2016.10a
• /
• pp.615-616
• /
• 2016

최근 4차 산업으로 패러다임이 변화함에 따라 SW산업이 더욱 중요하게 되었다. 이에 따라 전 세계적으로 코딩 교육에 대한 수요도 증가하게 되었고 기업에서도 SW를 잘 만들기 위한 코드 관리 중요성도 증가하게 되었다. 많은 양의 프로그래밍 소스코드를 사람이 일일이 채점하고 관리하는 것은 사실상 불가능하기 때문에 이러한 문제를 해결할 수 있는 코드 평가 시스템이 요구되고 있다. 하지만 어떤 코드가 좋은 코드인지 코드를 어떻게 평가해야하는지에 대한 명확한 기준은 없으며 이에 대한 연구도 부족한 상황이다. 최근에 주목 받고 있는 Deep Learning 기술은 이미지 처리, 자연어 처리등 기존의 Machine Learning 알고리즘이 냈던 성과보다 훨씬 뛰어난 성과를 내고 있다. 하지만 Programming language 영역에서는 아직 깊이 연구된 바가 없다. 따라서 본 연구에서는 Deep Learning 기술로 알려진 Convolutional Neural Network의 변형된 형태엔 Tree-based Convolutional Neural Network를 사용하여 프로그래밍 소스코드를 분석, 분류하는 알고리즘 및 코드의 Representation Learning에 대한 연구를 진행함으로써 이러한 문제를 해결하고자 한다.

• Wind and Structures
• /
• v.9 no.1
• /
• pp.37-58
• /
• 2006

The paper describes a study about effects of upstream hills on design wind loads using two mathematical approaches: Computational Fluid Dynamics (CFD) and Artificial Neural Network (NN for short). For this purpose CFD and NN tools have been developed using an object-oriented approach and C++ programming language. The CFD tool consists of solving the Reynolds time-averaged Navier-Stokes equations and $k-{\varepsilon}$ turbulence model using body-fitted nearly-orthogonal coordinate system. Subsequently, design wind load parameters such as speed-up ratio values have been generated for a wide spectrum of two-dimensional hill geometries that includes isolated and multiple steep and shallow hills. Ground roughness effect has also been considered. Such CFD solutions, however, normally require among other things ample computational time, background knowledge and high-capacity hardware. To assist the enduser, an easier, faster and more inexpensive NN model trained with the CFD-generated data is proposed in this paper. Prior to using the CFD data for training purposes, extensive validation work has been carried out by comparing with boundary layer wind tunnel (BLWT) data. The CFD trained NN (CFD-NN) has produced speed-up ratio values for cases such as multiple hills that are not covered by wind design standards such as the Commentaries of the National Building Code of Canada (1995). The CFD-NN results compare well with BLWT data available in literature and the proposed approach requires fewer resources compared to running BLWT experiments.

• Proceedings of the Korean Society for Agricultural Machinery Conference
• /
• 2000.11b
• /
• pp.124-131
• /
• 2000

In this study, the NEURAL NETWORK (hereinafter referred to as NN) was applied to control of the nonlinear factors for turning movement of the crawler vehicle and experiment was carried out using a small model of crawler vehicle in order to inspect an application of NN. Furthermore, CHAOS NEURAL NETWORK (hereinafter referred to as CNN) was also applied to this control so as to compare with conventional NN. CNN is especially effective for plane in many variables with local minimum which conventional NN is apt to fall into, and it is relatively useful to nonlinear factors. Experiment of turning on the slope of crawler vehicle was performed in order to estimate an adaptability of nonlinear problems by NN and CNN. The inclination angles of the road surface which the vehicles travel on, were respectively 4deg, 8deg, 12deg. These field conditions were selected by the object for changing nonlinear magnitude in turning phenomenon of vehicle. Learning of NN and CNN was carried out by referring to positioning data obtained from measurement at every 15deg in turning. After learning, the sampling data at every 15deg were interpolated based on the constructed learning system of NN and CNN. Learning and simulation programs of NN and CNN were made by C language ("Association of research for algorithm of calculating machine (1992)"). As a result, conventional NN and CNN were available for interpolation of sampling data. Moreover, when nonlinear intensity is not so large under the field condition of small slope, interpolation performance of CNN was a little not so better than NN. However, when nonlinear intensity is large under the field condition of large slope, interpolation performance of CNN was relatively better than NN.