• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.41 no.5
• /
• pp.592-601
• /
• 2016

The recurrent neural network (RNN) is a deep learning model which is suitable to sequential or length-variable data. The Long Short-Term Memory (LSTM) mitigates the vanishing gradient problem of RNNs so that LSTM can maintain the long-term dependency among the constituents of the given input sequence. In this paper, we propose a LSTM based language model which can predict following words of a given incomplete sentence to generate a complete sentence. To evaluate our method, we trained our model using multiple Korean corpora then generated the incomplete part of Korean sentences. The result shows that our language model was able to generate the fluent Korean sentences. We also show that the word based model generated better sentences compared to the other settings.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.17 no.7
• /
• pp.945-951
• /
• 2007

Noise cancellation is essential for the devices which use speech as an interface. In real environments, speech quality and recognition rates are degraded by the auditive noises coming near the microphone. In this paper, we propose a noise cancellation algorithm using stereo microphones basically. The advantage of the use of multiple microphones is that the direction information of the target source could be applied. The proposed noise canceller is based on the Wiener filter. To estimate the filter, noise and target speech frequency responses should be known and they are estimated by the spectral classification in the frequency domain. The performance of the proposed algorithm is compared with that of the well-known Frost algorithm and the generalized sidelobe canceller (GSC) with an adaptation mode controller (AMC). As performance measures, the perceptual evaluation of speech quality (PESQ), which is the most widely used among various objective speech quality methods, and speech recognition rates are adopted.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.12 no.1
• /
• pp.211-216
• /
• 2012

Control Theory for continuous-time system has been well developed. Due to the development of computer technology, digital control scheme are employed in many areas. When delays are in control systems, it is hard to control the system efficiently. Delays by controller-to-actuator and sensor-to-controller deteriorate control performance and could possibly destabilize the overall system. In this paper, a new approximated discretization method and digital design for control systems with multiple state, input and output delays and a generalized bilinear transformation method with a tunable parameter are also provided, which can re-transform the integer time-delayed discrete-time model to its continuous-time model. Illustrative example is given to demonstrate the effectiveness of the developed method.

• Smart Structures and Systems
• /
• v.4 no.6
• /
• pp.727-740
• /
• 2008

A down hole vertical seismic array is a sequence of instruments installed at various depths in the earth to record the ground motion at multiple points during an earthquake. Numerous studies demonstrate the unique utility of vertical seismic arrays for studying in situ site response and soil behavior. Examples are given of analyses made at two sites to show the value of data from vertical seismic arrays. The sites examined are the Lotung, Taiwan SMART1 array and a new site installed at Jingliao, Taiwan. Details of the installation of the Jingliao array are given. ARX models are theoretically the correct process models for vertical wave propagation in the layered earth, and are used to linearly map deeper sensor input signals to shallower sensor output signals. An example of Event 16 at the Lotung array is given. This same data, when examined in detail with a Bayesian inference model, can also be explained by nonlinear filters yielding commonly accepted soil degradation curves. Results from applying an ARMAX model to data from the Jingliao vertical seismic array are presented. Estimates of inter-transducer soil increment resonant frequency, shear modulus, and damping ratio are presented. The shear modulus varied from 50 to 150 MPa, and damping ratio between 8% and 15%. A new hardware monitoring system - TerraScope - is an affordable 4-D down-hole seismic monitoring system based on independent, microprocessor-controlled sensor Pods. The Pods are nominally 50 mm in diameter, and about 120 mm long. An internal 16-bit micro-controller oversees all aspects of instrumentation, eight programmable gain amplifiers, and local signal storage.

• The KIPS Transactions:PartA
• /
• v.12A no.2 s.92
• /
• pp.87-94
• /
• 2005

Pattern matching is one of critical parts of Network Intrusion Prevention Systems (NIPS) and computationally intensive. To handle a large number of attack signature fattens increasing everyday, a network intrusion prevention system requires a multi pattern matching method that can meet the line speed of packet transfer. In this paper, we analyze Snort, a widely used open source network intrusion prevention/detection system, and its pattern matching characteristics. A multi pattern matching method for NIPS should efficiently handle a large number of patterns with a wide range of pattern lengths and case insensitive patterns matches. It should also be able to process multiple input characters in parallel. We propose a multi pattern matching hardware accelerator based on Shift-OR pattern matching algorithm. We evaluate the performance of the pattern matching accelerator under various assumptions. The performance evaluation shows that the pattern matching accelerator can be more than 80 times faster than the fastest software multi-pattern matching method used in Snort.

• 제어로봇시스템학회:학술대회논문집
• /
• 2004.08a
• /
• pp.620-626
• /
• 2004

For general nonlinear processes, it is difficult to control with a linear model-based control method and nonlinear controls are considered. Among the numerous approaches suggested, the most rigorous approach is to use dynamic optimization. Many general engineering problems like control, scheduling, planning etc. are expressed by functional optimization problem and most of them can be changed into dynamic programming (DP) problems. However the DP problems are used in just few cases because as the size of the problem grows, the dynamic programming approach is suffered from the burden of calculation which is called as 'curse of dimensionality'. In order to avoid this problem, the Neuro-Dynamic Programming (NDP) approach is proposed by Bertsekas and Tsitsiklis (1996). To get the solution of seriously nonlinear process control, the interest in NDP approach is enlarged and NDP algorithm is applied to diverse areas such as retailing, finance, inventory management, communication networks, etc. and it has been extended to chemical engineering parts. In the NDP approach, we select the optimal control input policy to minimize the value of cost which is calculated by the sum of current stage cost and future stages cost starting from the next state. The cost value is related with a weight square sum of error and input movement. During the calculation of optimal input policy, if the approximate cost function by using simulation data is utilized with Bellman iteration, the burden of calculation can be relieved and the curse of dimensionality problem of DP can be overcome. It is very important issue how to construct the cost-to-go function which has a good approximate performance. The neural network is one of the eager learning methods and it works as a global approximator to cost-to-go function. In this algorithm, the training of neural network is important and difficult part, and it gives significant effect on the performance of control. To avoid the difficulty in neural network training, the lazy learning method like k-nearest neighbor method can be exploited. The training is unnecessary for this method but requires more computation time and greater data storage. The pH neutralization process has long been taken as a representative benchmark problem of nonlin ar chemical process control due to its nonlinearity and time-varying nature. In this study, the NDP algorithm was applied to pH neutralization process. At first, the pH neutralization process control to use NDP algorithm was performed through simulations with various approximators. The global and local approximators are used for NDP calculation. After that, the verification of NDP in real system was made by pH neutralization experiment. The control results by NDP algorithm was compared with those by the PI controller which is traditionally used, in both simulations and experiments. From the comparison of results, the control by NDP algorithm showed faster and better control performance than PI controller. In addition to that, the control by NDP algorithm showed the good results when it applied to the cases with disturbances and multiple set point changes.

• Journal of Intelligence and Information Systems
• /
• v.9 no.1
• /
• pp.227-249
• /
• 2003

Prediction of corporate failure using past financial data is a well-documented topic. Early studies of bankruptcy prediction used statistical techniques such as multiple discriminant analysis, logit and probit. Recently, however, numerous studies have demonstrated that artificial intelligence such as neural networks can be an alternative methodology for classification problems to which traditional statistical methods have long been applied. In building neural network model, the selection of independent and dependent variables should be approached with great care and should be treated as model construction process. Irrespective of the efficiency of a teaming procedure in terms of convergence, generalization and stability, the ultimate performance of the estimator will depend on the relevance of the selected input variables and the quality of the data used. Approaches developed in statistical methods such as correlation analysis and stepwise selection method are often very useful. These methods, however, may not be the optimal ones for the development of neural network model. In this paper, we propose a genetic algorithms approach to find an optimal or near optimal input variables fur neural network modeling. The proposed approach is demonstrated by applications to bankruptcy prediction modeling. Our experimental results show that this approach increases overall classification accuracy rate significantly.

• Journal of KIISE:Computer Systems and Theory
• /
• v.32 no.11_12
• /
• pp.629-638
• /
• 2005

In the System on a Chip design, the on chip bus is one of the critical factors that decides the overall system performance. Especially, in the case or reusing the IPs such as processors, DSPs and multimedia IPs that requires higher bandwidth, the bandwidth problems of on chip bus are getting more serious. Recently ARM proposes the Multi-Layer AHB BusMatrix that is a highly efficient on chip bus to solve the bandwidth problems. The Multi-Layer AHB BusMatrix allows parallel access paths between multiple masters and slaves in a system. This is achieved by using a more complex interconnection matrix and gives the benefit of increased overall bus bandwidth, and a more flexible system architecture. However, there is one clock cycle delay for each master in existing Multi-Layer AHB BusMatrix whenever the master starts new transactions or changes the slave layers because of the Input Stage and arbitration logic realized with Moore type. In this paper, we improved the existing Multi-Layer AHB BusMatrix architecture to solve the one clock cycle delay problems and to reduce the area overhead of the Input Stage. With the elimination of the Input Stage and some restrictions on the arbitration scheme, we tan take away the one clock cycle delay and reduce the area overhead. Experimental results show that the end time of total bus transaction and the average latency time of improved Multi-Layer AHB BusMatrix are improved by $20\%\;and\;24\%$ respectively. in ease of executing a number of transactions by 4-beat incrementing burst type. Besides the total area and the clock period are reduced by $22\%\;and\;29\%$ respectively, compared with existing Multi-layer AHB BusMatrix.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.3
• /
• pp.319-325
• /
• 2022

The variation of utterance lengths is a representative factor that can degrade the performance of speaker verification systems. To handle this issue, previous studies had attempted to extract speaker features from various branches or to use convolution layers with different receptive fields. Combining the advantages of the previous two approaches for variable-length input, this paper proposes integrated receptive field diversification that extracts speaker features through more diverse receptive field. The proposed method processes the input features by convolutional layers with different receptive fields at multiple time-axis branches, and extracts speaker embedding by dynamically aggregating the processed features according to the lengths of input utterances. The deep neural networks in this study were trained on the VoxCeleb2 dataset and tested on the VoxCeleb1 evaluation dataset that divided into 1 s, 2 s, 5 s, and full-length. Experimental results demonstrated that the proposed method reduces the equal error rate by 19.7 % compared to the baseline.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.21 no.3
• /
• pp.523-529
• /
• 2017

In this paper, we propose an efficient sphere decoding scheme that reduces computational complexity by combining receive and transmit ordering techniques in generalized spatial modulation systems, where the indexes of activated transmit antennas as well as the transmit symbols are exploited to transfer information to the receiver. In this scheme, the receive signals are optimally ordered so that the calculation for a candidate solution outside the sphere is terminated early to lower the computational complexity. In addition, the transmit ordering technique is applied to first search for candidate symbols and activated antennas having higher probabilities to further reduce the computational complexity. Simulation results show that the proposed doubly ordered sphere decoding scheme provides the same bit error rate performance with the conventional sphere decoding method and the sphere decoder employing only the receive ordering technique while it requires lower computational complexity.