• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.15-23
• /
• 2001

This paper presents a voice conversion technique that modifies the utterance of a source speaker as if it were spoken by a target speaker. Feature parameter conversion methods to perform the transformation of vocal tract and prosodic characteristics between the source and target speakers are described. The transformation of vocal tract characteristics is achieved by modifying the LPC cepstral coefficients using Linear Multivariate Regression (LMR). Prosodic transformation is done by changing the average pitch period between speakers, and it is applied to the residual signal using the LP-PSOLA scheme. Experimental results show that transformed speech by LMR and LP-PSOLA synthesis method contains much characteristics of the target speaker.

• The KIPS Transactions:PartD
• /
• v.11D no.1
• /
• pp.163-172
• /
• 2004

In the software development project applying object-oriented development methodology, the research on the UCP(Use Case Point) as a method to estimate development effort is being carried on. The existing research proposes the linear model calculating the development effort that multiplies an invariant on AUCP(Adjusted Use Case Point) which applied technical and environmental factors. However, the statistical model that estimates the development effort using AUCP and UUCP(Unadjusted Use Case Point) is not being studied. The irrelevant relationship of the linear regression model, whose development period is increasing tremendously as the software size increases, is confirmed. Moreover, during the UCP calculating process, there can be errors in FP by applying the TCF(Technical Complexity Factor) and EF(Environmental Factor). This paper presents a non-linear regression model, that does not consider the TCF and EF, and that estimate the development effort from UUCP directly by utilizing the exponential function. An exponential function is selected among the linear, logarithm, polynomial, power, and exponential model via statistical evaluations of the models mentioned above.

• Journal of Broadcast Engineering
• /
• v.26 no.2
• /
• pp.184-196
• /
• 2021

This paper proposes statistics adaptive linear regression-based object size prediction method for object detection. YOLOv2 and YOLOv3, which are typical deep learning-based object detection algorithms, designed the last layer of a network using statistics adaptive exponential regression model to predict the size of objects. However, an exponential regression model can propagate a high derivative of a loss function into all parameters in a network because of the property of an exponential function. We propose statistics adaptive linear regression layer to ease the gradient exploding problem of the exponential regression model. The proposed statistics adaptive linear regression model is used in the last layer of the network to predict the size of objects with statistics estimated from training dataset. We newly designed the network based on the YOLOv3tiny and it shows the higher performance compared to YOLOv3 tiny on the UFPR-ALPR dataset.

• 한국정보교육학회:학술대회논문집
• /
• 2021.08a
• /
• pp.51-57
• /
• 2021

Quite a few elementary school teachers began to utilize AI technology in order to provide students with customized, intelligent information services in recent years. However, learning principles of AI may be as important as utilizing AI in everyday life because understanding principles of AI can empower them to buildup adaptability to changes in highly technological world. In the paper, 'Linear Regression Algorithm' is selected for teaching AI-based prediction system to solve real world problems suitable for elementary students. A simulation program written in Scratch was developed so that students can find a solution of linear regression model using the program. The paper shows that students have learned analyzing data as well as comparing the accuracy of the prediction model. Also, they have shown the ability to solve real world problems by finding suitable prediction models.

• The Korean Journal of Applied Statistics
• /
• v.31 no.1
• /
• pp.123-137
• /
• 2018

A nonlinear mixed effects model is mainly used to analyze repeated measurement data in various fields. A nonlinear mixed effects model consists of two stages: the first-stage individual-level model considers intra-individual variation and the second-stage population model considers inter-individual variation. The individual-level model, which is the first stage of the nonlinear mixed effects model, estimates the parameters of the nonlinear regression model. It is the same as the general nonlinear regression model, and usually estimates parameters using the least squares estimation method. However, the least squares estimation method may have a problem that the estimated value of the parameters and standard errors become extremely large if the assumed nonlinear function is not explicitly revealed by the data. In this paper, a new estimation method is proposed to solve this problem by introducing the ridge regression method recently proposed in the nonlinear regression model into the first-stage individual-level model of the nonlinear mixed effects model. The performance of the proposed estimator is compared with the performance with the standard estimator through a simulation study. The proposed methodology is also illustrated using quantitative high throughput screening data obtained from the US National Toxicology Program.

• The Korean Journal of Applied Statistics
• /
• v.33 no.1
• /
• pp.1-10
• /
• 2020

We propose a unified approach to variable selection, transformation and outliers in the linear model. The procedure includes a sequential method for outlier detection and a least trimmed squares estimator for variable transformation. It uses all possible subsets regressions for model selection. Some real data analyses and the simulation results are provided to show the efficiency of the methods in the context of the correct variable selection and the fitness of the estimated model.

• Journal of Convergence for Information Technology
• /
• v.9 no.6
• /
• pp.1-6
• /
• 2019

In this study, the development of a weight estimation model of electronic scale with nonlinear characteristics is presented using polynomial regression analysis. The output voltage of the load cell was measured directly using the reference mass. And a polynomial regression model was obtained using the matrix and curve fitting function of MS Office Excel. The weight was measured in 100g units using a load cell electronic scale measuring up to 5kg and the polynomial regression model was obtained. The error was calculated for simple($1^{st}$), $2^{nd}$ and $3^{rd}$ order polynomial regression. To analyze the suitability of the regression function for each model, the coefficient of determination was presented to indicate the correlation between the estimated mass and the measured data. Using the third order polynomial model proposed here, a very accurate model was obtained with a standard deviation of 10g and the determinant coefficient of 1.0. Based on the theory of multi regression model presented here, it can be used in various statistical researches such as weather forecast, new drug development and economic indicators analysis using logistic regression analysis, which has been widely used in artificial intelligence fields.

• Journal of the Korea Society for Simulation
• /
• v.18 no.4
• /
• pp.137-148
• /
• 2009

The purpose of this study is to implement variable selection algorithm which helps construct a reliable linear regression model. If we use all candidate variables to construct a linear regression model, the significance of the model will be decreased and it will cause 'Curse of Dimensionality'. And if the number of data is less than the number of variables (dimension), we cannot construct the regression model. Due to these problems, we consider the variable selection problem as a combinatorial optimization problem, and apply GA (Genetic Algorithm) to the problem. Typical measures of estimating statistical significance are $R^2$, F-value of regression model, t-value of regression coefficients, and standard error of estimates. We design GA to solve multi-objective functions, because statistical significance of model is not to be estimated by a single measure. We perform experiments using simulation data, designed to consider various kinds of situations. As a result, it shows better performance than LARS (Least Angle Regression) which is an algorithm to solve variable selection problems. We modify algorithm to solve portfolio selection problem which construct portfolio by selecting stocks. We conclude that the algorithm is able to solve real problems.

• Proceedings of the Korean Information Science Society Conference
• /
• 2002.10c
• /
• pp.694-696
• /
• 2002

이 논문에서는 시스템 수준의 전력 소모를 분석하는 방법론을 설명한다. 응용의 시스템 수준 전력 모델을 구하기 위해서, 시스템을 이루는 각 부분들을 선형적으로 모델링하고, 이를 모두 더한다. 선형적으로 모델링된 식의 파라메터들을 구하기 위해서, 회귀분석에 기반한 분석을 한다. 이를 위해서 다양한 벤치마크들을 준비하고, 응용에 대해서 측정을 한 것과 수정된 시뮬레이터에서 필요한 정보를 얻어야 한다. 이렇게 분석한 전력 모델의 예측치는 5% 내의 정확도를 가짐을 확인하였다.

• Transactions of the Korean Society of Mechanical Engineers B
• /
• v.35 no.11
• /
• pp.1111-1117
• /
• 2011

This paper proposed a fault detection and diagnosis (FDD) algorithm using nonlinear regression models, focusing especially on heat exchanger faults. This research concerned four working modes: those with no fault, evaporator fault, condenser fault, and evaporator and condenser faults. This research used no fault mode data to create an FDD algorithm. Using the no fault mode data, correlation functions for predicting the degree of superheat or subcool of heat exchangers (an evaporator and a condenser) were derived. Each correlation function has five inputs and one output. Based on these correlation functions, it is possible to predict the degree of superheat or subcool of each heat exchanger under various working conditions. The FDD algorithm was developed by comparing the predicted value and the simulation value. The FDD algorithm works well in all four working modes.